wchess : minor

* Add first draft of server

* Added json support and base funcs for server.cpp

* Add more user input via api-request

also some clean up

* Add reqest params and load post function

Also some general clean up

* Remove unused function

* Add readme

* Add exception handlers

* Update examples/server/server.cpp

* make : add server target

* Add magic curl syntax

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.devops/cublas.Dockerfile

@@ -0,0 +1,28 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV WHISPER_CUBLAS=1
+
+RUN make
+
+ENTRYPOINT ["/app/main"]

.github/workflows/build.yml

@@ -1,31 +1,41 @@
 name: CI
 on: [push, pull_request]
 
+env:
+  ubuntu_image: "ubuntu:22.04"
+
 jobs:
   ubuntu-latest:
     runs-on: ubuntu-latest
 
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install libsdl2-dev
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
 
-      - name: Build
+      - name: Build ${{ matrix.arch }}
         run: |
-          make
-          make stream
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential libsdl2-dev
+            make
+            make stream'
 
   macOS-latest:
     runs-on: macOS-latest
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Dependencies
         run: |
@@ -37,82 +47,104 @@ jobs:
           make
           make stream
 
+  freeBSD-latest:
+    runs-on: macos-12
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Build
+        uses: cross-platform-actions/action@v0.15.0
+        with:
+          operating_system: freebsd
+          version: '13.2'
+          run: |
+            sudo pkg update
+            sudo pkg install -y gmake sdl2
+            gmake
+            gmake stream
+
   ubuntu-latest-gcc:
     runs-on: ubuntu-latest
 
     strategy:
+      fail-fast: false
       matrix:
         build: [Debug, Release]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-          sudo apt-get install libsdl2-dev
-
-      - name: Configure
-        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+            make
+            ctest -L gh --output-on-failure'
 
   ubuntu-latest-clang:
     runs-on: ubuntu-latest
 
     strategy:
+      fail-fast: false
       matrix:
         build: [Debug, Release]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-          sudo apt-get install libsdl2-dev
-
-      - name: Configure
-        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+            make
+            ctest -L gh --output-on-failure'
 
   ubuntu-latest-gcc-sanitized:
     runs-on: ubuntu-latest
 
     strategy:
+      fail-fast: false
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        arch: [linux/amd64]
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v2
+
+      - name: Build ${{ matrix.arch }}
         run: |
-          sudo apt-get update
-          sudo apt-get install build-essential
-          sudo apt-get install cmake
-
-      - name: Configure
-        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
-
-      - name: Build
-        run: |
-          make
-          ctest -L gh --output-on-failure
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential cmake
+            cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
+            make
+            ctest -L gh --output-on-failure'
 
   windows:
     runs-on: windows-latest
@@ -125,14 +157,16 @@ jobs:
         include:
           - arch: Win32
             s2arc: x86
+            jnaPath: win32-x86
           - arch: x64
             s2arc: x64
+            jnaPath: win32-x86-64
           - sdl2: ON
             s2ver: 2.26.0
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
@@ -148,7 +182,7 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
         run: |
@@ -159,6 +193,12 @@ jobs:
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
+      - name: Upload dll
+        uses: actions/upload-artifact@v3
+        with:
+          name: ${{ matrix.jnaPath }}_whisper.dll
+          path: build/bin/${{ matrix.build }}/whisper.dll
+
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
@@ -177,17 +217,17 @@ jobs:
         sdl2: [ON]
         include:
           - arch: Win32
-            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-x86.zip
             s2arc: x86
           - arch: x64
-            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
+            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.24/OpenBLAS-0.3.24-x64.zip
             s2arc: x64
           - sdl2: ON
             s2ver: 2.26.0
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
@@ -199,7 +239,7 @@ jobs:
           7z x blas.zip -oblas -y
           copy blas/include/cblas.h .
           copy blas/include/openblas_config.h .
-          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+          echo "OPENBLAS_PATH=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
 
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
@@ -212,9 +252,9 @@ jobs:
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_SUPPORT_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
-          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_OPENBLAS=${{ matrix.blas }}
+          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
+          -DWHISPER_SDL2=${{ matrix.sdl2 }}
 
       - name: Build
         run: |
@@ -223,7 +263,7 @@ jobs:
 
       - name: Copy libopenblas.dll
         if: matrix.blas == 'ON'
-        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+        run: copy "$env:OPENBLAS_PATH/bin/libopenblas.dll" build/bin/${{ matrix.build }}
 
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
@@ -235,10 +275,10 @@ jobs:
         with:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
-          
+
   windows-cublas:
     runs-on: windows-latest
-    
+
     strategy:
       matrix:
         build: [Release]
@@ -250,40 +290,47 @@ jobs:
             s2arc: x64
           - sdl2: ON
             s2ver: 2.26.0
-            
+
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
-        
+
       - name: Install CUDA Toolkit
         id: cuda-toolkit
         uses: Jimver/cuda-toolkit@v0.2.10
-        
+
       - name: Fetch SDL2 and set SDL2_DIR
         if: matrix.sdl2 == 'ON'
         run: |
           C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
-        
+
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           -DWHISPER_CUBLAS=1
-      
+
       - name: Build
         run: |
           cd ./build
           msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
-          
+
+      - name: Copy CUDA DLLs
+        run: >
+          Copy-Item -PassThru
+          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
+          -Include cudart64_*,cublas64_*,cublasLt64_*
+          -Destination build/bin/${{ matrix.build }}
+
       - name: Copy SDL2.dll
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
-          
+
       - name: Upload binaries
         if: matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
@@ -300,24 +347,16 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
-      - name: Dependencies
-        run: |
-          wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
-          tar -xvf master.tar.gz
-          emsdk-master/emsdk update
-          emsdk-master/emsdk install latest
-          emsdk-master/emsdk activate latest
+      - name: Setup emsdk
+        uses: mymindstorm/setup-emsdk@v12
 
-      - name: Configure
-        run: echo "tmp"
+      - name: Verify
+        run: emcc -v
 
       - name: Build
         run: |
-          pushd emsdk-master
-          source ./emsdk_env.sh
-          popd
           emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           make
 
@@ -330,7 +369,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Configure
         run: |
@@ -339,7 +378,7 @@ jobs:
 
       - name: Build objc example
         run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
-      
+
       - name: Build swiftui example
         run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
 
@@ -348,14 +387,14 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Install Java
         uses: actions/setup-java@v3
         with:
           distribution: zulu
           java-version: 17
-      
+
       - name: Setup Android SDK
         uses: android-actions/setup-android@v2
 
@@ -363,3 +402,84 @@ jobs:
         run: |
           cd examples/whisper.android
           ./gradlew assembleRelease --no-daemon
+
+  android_java:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: set up JDK 11
+        uses: actions/setup-java@v3
+        with:
+          java-version: '11'
+          distribution: 'temurin'
+          cache: gradle
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v2
+        with:
+          api-level: 30
+          build-tools-version: 30.0.3
+
+      - name: Build
+        run: |
+          cd examples/whisper.android.java
+          chmod +x ./gradlew 
+          ./gradlew assembleRelease
+
+  java:
+    needs: [ 'windows' ]
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Install Java
+        uses: actions/setup-java@v1
+        with:
+          java-version: 17
+
+      - name: Download Windows lib
+        uses: actions/download-artifact@v3
+        with:
+          name: win32-x86-64_whisper.dll
+          path: bindings/java/build/generated/resources/main/win32-x86-64
+
+      - name: Build
+        run: |
+          models\download-ggml-model.cmd tiny.en
+          cd bindings/java
+          chmod +x ./gradlew
+          ./gradlew build
+
+      - name: Upload jar
+        uses: actions/upload-artifact@v3
+        with:
+          name: whispercpp.jar
+          path: bindings/java/build/libs/whispercpp-*.jar
+
+      - name: Publish package
+        if: ${{ github.ref == 'refs/heads/master' }}
+        uses: gradle/gradle-build-action@v2.4.2
+        with:
+          arguments: publish
+          build-root-directory: bindings/java
+        env:
+          MAVEN_USERNAME: ${{ secrets.JIRA_USER }}
+          MAVEN_PASSWORD: ${{ secrets.JIRA_PASS }}
+          PGP_SECRET: ${{ secrets.GPG_PRIVATE_KEY }}
+          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+
+  quantize:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Test quantize
+        run: |
+          ./models/download-ggml-model.sh tiny.en
+          make quantize
+          ./quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0

.gitignore

@@ -8,15 +8,22 @@
 .DS_Store
 
 build/
+build-coreml/
 build-em/
 build-debug/
 build-release/
+build-rwdi/
 build-static/
 build-cublas/
 build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
+# SPM
+.build/
+.swiftpm
+*.metallib
+
 /main
 /stream
 /command
@@ -24,6 +31,8 @@ build-sanitize-thread/
 /talk-llama
 /bench
 /quantize
+/server
+/lsp
 
 arm_neon.h
 sync.sh
@@ -41,3 +50,12 @@ extra/bench-gg.txt
 models/*.mlmodel
 models/*.mlmodelc
 models/*.mlpackage
+bindings/java/.gradle/
+bindings/java/.idea/
+.idea/
+
+benchmark_results.csv
+cmake-build-debug/
+.cxx/
+.gradle/
+local.properties

CMakeLists.txt

@@ -1,6 +1,6 @@
-cmake_minimum_required (VERSION 3.0)
+cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.4.2)
+project(whisper.cpp VERSION 1.5.0)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -35,6 +35,12 @@ endif()
 
 # options
 
+if (APPLE)
+    set(WHISPER_METAL_DEFAULT ON)
+else()
+    set(WHISPER_METAL_DEFAULT OFF)
+endif()
+
 option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
 
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
@@ -54,14 +60,21 @@ option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
 option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
 option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)
 
+option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
+
 if (APPLE)
     option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
+    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
+    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
     option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
     option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
 else()
-    option(WHISPER_OPENBLAS              "whisper: support for OpenBLAS" OFF)
-    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"   OFF)
-    option(WHISPER_CLBLAST               "whisper: use CLBlast"          OFF)
+    option(WHISPER_BLAS                  "whisper: use BLAS libraries"  OFF)
+    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor" Generic)
+    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"     OFF)
+    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"  OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS" OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
 endif()
 
 option(WHISPER_PERF "whisper: enable perf timings" OFF)
@@ -104,10 +117,38 @@ if (APPLE)
             set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
         else()
-            message(WARNING "Accelerate framework not found")
+            message(FATAL_ERROR "Accelerate framework not found")
         endif()
     endif()
 
+    if (WHISPER_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+        if (METAL_FRAMEWORK)
+            message(STATUS "Metal framework found")
+
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
+                ${FOUNDATION_LIBRARY}
+                ${METAL_FRAMEWORK}
+                ${METALKIT_FRAMEWORK}
+                )
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
+
+            if (WHISPER_METAL_NDEBUG)
+                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
+            endif()
+        else()
+            message(FATAL_ERROR "Metal framework not found")
+        endif()
+
+        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+        # copy ggml-metal.metal to bin directory
+        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+    endif()
+
     if (WHISPER_COREML)
         find_library(FOUNDATION_FRAMEWORK Foundation)
         find_library(COREML_FRAMEWORK CoreML)
@@ -117,7 +158,7 @@ if (APPLE)
 
             set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
         else()
-            message(WARNING "CoreML framework not found")
+            message(FATAL_ERROR "CoreML framework not found")
         endif()
 
         if (WHISPER_COREML_ALLOW_FALLBACK)
@@ -127,19 +168,41 @@ if (APPLE)
 endif()
 
 if (WHISPER_OPENBLAS)
-    find_library(OPENBLAS_LIB
-        NAMES openblas libopenblas
-        )
-    if (OPENBLAS_LIB)
-        message(STATUS "OpenBLAS found")
-
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${OPENBLAS_LIB})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
-    else()
-        message(WARNING "OpenBLAS not found")
-    endif()
+    set(WHISPER_BLAS_VENDOR "OpenBLAS")
+    set(WHISPER_BLAS ON)
 endif()
 
+if (WHISPER_BLAS)
+    if (WIN32)
+        if(DEFINED ENV{OPENBLAS_PATH})
+            set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories($ENV{OPENBLAS_PATH}/include)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else ()
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
+        endif ()
+    else ()
+        set(BLA_STATIC 1)
+        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
+        set(BLA_SIZEOF_INTEGER 8)
+        set(BLA_PREFER_PKGCONFIG 1)
+        find_package(BLAS)
+
+        if(BLAS_FOUND)
+            message(STATUS "BLAS compatible library found")
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories(${BLAS_INCLUDE_DIRS})
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else()
+            message(FATAL_ERROR "BLAS library was not found")
+        endif()
+    endif ()
+endif ()
+
 if (WHISPER_CUBLAS)
     cmake_minimum_required(VERSION 3.17)
 
@@ -150,7 +213,7 @@ if (WHISPER_CUBLAS)
 
         enable_language(CUDA)
 
-        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
 
         add_compile_definitions(GGML_USE_CUBLAS)
 
@@ -161,7 +224,38 @@ if (WHISPER_CUBLAS)
         endif()
 
     else()
-        message(WARNING "cuBLAS not found")
+        message(FATAL_ERROR "cuBLAS not found")
+    endif()
+endif()
+
+
+if (WHISPER_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (WHISPER_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
     endif()
 endif()
 
@@ -170,16 +264,20 @@ if (WHISPER_CLBLAST)
     if (CLBlast_FOUND)
         message(STATUS "CLBlast found")
 
-        set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h)
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
 
         add_compile_definitions(GGML_USE_CLBLAST)
 
         set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
     else()
-        message(WARNING "CLBlast not found")
+        message(FATAL_ERROR "CLBlast not found")
     endif()
 endif()
 
+if( WHISPER_OPENVINO )
+    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+endif()
+
 # compiler flags
 
 if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
@@ -219,20 +317,25 @@ message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 
 if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
     message(STATUS "ARM detected")
+elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+    message(STATUS "PowerPC detected")
 else()
     message(STATUS "x86 detected")
     if (MSVC)
-            if(NOT WHISPER_NO_AVX2)
-                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
-                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
-                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
-            else()
-                if(NOT WHISPER_NO_AVX)
-                    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
-                    set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
-                    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
-                endif()
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
+        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
+        if(NOT WHISPER_NO_AVX2)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
+            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
+        else()
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
             endif()
+        endif()
     else()
         if (EMSCRIPTEN)
             set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
@@ -254,6 +357,53 @@ else()
     endif()
 endif()
 
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
 if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
@@ -285,6 +435,24 @@ if (WHISPER_COREML)
         )
 endif()
 
+if (WHISPER_OPENVINO)
+    set(TARGET whisper.openvino)
+
+    add_library(${TARGET} OBJECT
+        openvino/whisper-openvino-encoder.h
+        openvino/whisper-openvino-encoder.cpp
+        )
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
+
+    target_link_libraries(${TARGET} PRIVATE openvino::runtime)
+endif()
+
 #
 # whisper - this is the main library of the project
 #
@@ -294,8 +462,15 @@ set(TARGET whisper)
 add_library(${TARGET}
     ggml.h
     ggml.c
-    ${GGML_CUDA_SOURCES}
-    ${GGML_OPENCL_SOURCES}
+    ggml-alloc.h
+    ggml-alloc.c
+    ggml-backend.h
+    ggml-backend.c
+    ggml-quants.h
+    ggml-quants.c
+    ${GGML_SOURCES_METAL}
+    ${GGML_SOURCES_CUDA}
+    ${GGML_SOURCES_OPENCL}
     whisper.h
     whisper.cpp
     )
@@ -310,6 +485,10 @@ if (WHISPER_COREML)
     target_link_libraries(${TARGET} PRIVATE whisper.coreml)
 endif()
 
+if (WHISPER_OPENVINO)
+    target_link_libraries(${TARGET} PRIVATE whisper.openvino)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 
@@ -332,9 +511,15 @@ if (BUILD_SHARED_LIBS)
         WHISPER_BUILD
         GGML_BUILD
         )
+
+    if (WHISPER_METAL)
+        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
+        #       but for some reason it does not work here like it does in llama.cpp
+        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
 endif()
 
-if (GGML_CUDA_SOURCES)
+if (GGML_SOURCES_CUDA)
     message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
     set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES OFF)
     set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
@@ -350,10 +535,13 @@ target_compile_definitions(${TARGET} PUBLIC
 
 set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
 
+include(GNUInstallDirs)
+
 install(TARGETS ${TARGET}
-    LIBRARY DESTINATION lib
-    ARCHIVE DESTINATION lib/static
-    RUNTIME DESTINATION bin
+    LIBRARY  DESTINATION lib
+    ARCHIVE  DESTINATION lib/static
+    RUNTIME  DESTINATION bin
+    RESOURCE DESTINATION bin
     PUBLIC_HEADER DESTINATION include
     )
 

Makefile

@@ -1,4 +1,4 @@
-default: main bench quantize
+default: main bench quantize server
 
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@@ -12,7 +12,13 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
-CCV := $(shell $(CC) --version | head -n 1)
+ifndef NVCC_VERSION
+	ifeq ($(call,$(shell which nvcc))$(.SHELLSTATUS),0)
+		NVCC_VERSION := $(shell nvcc --version | egrep -o "V[0-9]+.[0-9]+.[0-9]+" | cut -c2-)
+	endif
+endif
+
+CCV  := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)
 
 # Mac OS + Arm can report x86_64
@@ -36,27 +42,59 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
-# ref: https://github.com/ggerganov/whisper.cpp/issues/37
-ifneq ($(wildcard /usr/include/musl/*),)
-	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
-	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+CFLAGS   += -D_XOPEN_SOURCE=600
+CXXFLAGS += -D_XOPEN_SOURCE=600
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+ifeq ($(UNAME_S),OpenBSD)
+	CFLAGS   += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+	CXXFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+endif
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+ifeq ($(UNAME_S),Linux)
+	CFLAGS   += -D_GNU_SOURCE
+	CXXFLAGS += -D_GNU_SOURCE
+endif
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+ifeq ($(UNAME_S),Darwin)
+	CFLAGS   += -D_DARWIN_C_SOURCE
+	CXXFLAGS += -D_DARWIN_C_SOURCE
+endif
+ifeq ($(UNAME_S),DragonFly)
+	CFLAGS   += -D__BSD_VISIBLE
+	CXXFLAGS += -D__BSD_VISIBLE
+endif
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+ifeq ($(UNAME_S),FreeBSD)
+	CFLAGS   += -D__BSD_VISIBLE
+	CXXFLAGS += -D__BSD_VISIBLE
+endif
+ifeq ($(UNAME_S),NetBSD)
+	CFLAGS   += -D_NETBSD_SOURCE
+	CXXFLAGS += -D_NETBSD_SOURCE
+endif
+ifeq ($(UNAME_S),OpenBSD)
+	CFLAGS   += -D_BSD_SOURCE
+	CXXFLAGS += -D_BSD_SOURCE
 endif
 
 # OS specific
 # TODO: support Windows
-ifeq ($(UNAME_S),Linux)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Darwin)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),FreeBSD)
-	CFLAGS   += -pthread
-	CXXFLAGS += -pthread
-endif
-ifeq ($(UNAME_S),Haiku)
+ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$(UNAME_S))
 	CFLAGS   += -pthread
 	CXXFLAGS += -pthread
 endif
@@ -64,66 +102,56 @@ endif
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	ifeq ($(UNAME_S),Darwin)
-		CFLAGS += -mf16c
-		AVX1_M := $(shell sysctl machdep.cpu.features)
-		ifneq (,$(findstring FMA,$(AVX1_M)))
-			CFLAGS += -mfma
-		endif
-		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysctl machdep.cpu.leaf7_features)
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
+		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
 	else ifeq ($(UNAME_S),Linux)
-		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell grep "fma " /proc/cpuinfo)
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-
-			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
-		endif
-		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
-		ifneq (,$(findstring sse3,$(SSE3_M)))
-			CFLAGS += -msse3
-		endif
+		CPUINFO_CMD := cat /proc/cpuinfo
+	else ifneq (,$(filter MINGW32_NT% MINGW64_NT%,$(UNAME_S)))
+		CPUINFO_CMD := cat /proc/cpuinfo
+	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
+		CPUINFO_CMD := grep Features /var/run/dmesg.boot
 	else ifeq ($(UNAME_S),Haiku)
-		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell sysinfo -cpu | grep "FMA ")
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-
-			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-			ifneq (,$(findstring avx,$(AVX1_M)))
-				CFLAGS += -mavx
-			endif
-		endif
-	else
-		CFLAGS += -mfma -mf16c -mavx -mavx2
+		CPUINFO_CMD := sysinfo -cpu
+	endif
+
+	ifdef CPUINFO_CMD
+		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
+		ifneq (,$(AVX_M))
+			CFLAGS   += -mavx
+			CXXFLAGS += -mavx
+		endif
+
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
+		ifneq (,$(AVX2_M))
+			CFLAGS   += -mavx2
+			CXXFLAGS += -mavx2
+		endif
+
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
+		endif
+
+		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
+		ifneq (,$(F16C_M))
+			CFLAGS   += -mf16c
+			CXXFLAGS += -mf16c
+		endif
+
+		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
+		ifneq (,$(SSE3_M))
+			CFLAGS   += -msse3
+			CXXFLAGS += -msse3
+		endif
+
+		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
+		ifneq (,$(SSSE3_M))
+			CFLAGS   += -mssse3
+			CXXFLAGS += -mssse3
+		endif
 	endif
-endif
-ifeq ($(UNAME_M),amd64)
-	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
 
 ifneq ($(filter ppc64%,$(UNAME_M)),)
@@ -154,30 +182,67 @@ ifdef WHISPER_COREML_ALLOW_FALLBACK
 endif
 endif
 
+ifndef WHISPER_NO_METAL
+	ifeq ($(UNAME_S),Darwin)
+		WHISPER_METAL := 1
+
+		CFLAGS   += -DGGML_USE_METAL
+		CXXFLAGS += -DGGML_USE_METAL
+		LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	endif
+endif
+
 ifdef WHISPER_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas
 endif
 
 ifdef WHISPER_CUBLAS
-	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
+		CUDA_ARCH_FLAG=native
+	else
+		CUDA_ARCH_FLAG=all
+	endif
+
+	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
 	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
-	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=native
+	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif
 
+ifdef WHISPER_HIPBLAS
+	ROCM_PATH   ?= /opt/rocm
+	HIPCC       ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	CFLAGS      += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	CXXFLAGS    += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	LDFLAGS     += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	LDFLAGS     += -lhipblas -lamdhip64 -lrocblas
+	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	WHISPER_OBJ += ggml-cuda.o
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif
+
 ifdef WHISPER_CLBLAST
 	CFLAGS 		+= -DGGML_USE_CLBLAST
-	LDFLAGS	 	+= -lclblast -lOpenCL
+	CXXFLAGS 	+= -DGGML_USE_CLBLAST
+	LDFLAGS	 	+= -lclblast
+	ifeq ($(UNAME_S),Darwin)
+		LDFLAGS	 	+= -framework OpenCL
+	else
+		LDFLAGS	    += -lOpenCL
+	endif
 	WHISPER_OBJ	+= ggml-opencl.o
-	
-ggml-opencl.o: ggml-opencl.c ggml-opencl.h
-	$(CC) $(CFLAGS) -c $< -o $@
+
+ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif
 
 ifdef WHISPER_GPROF
@@ -233,6 +298,17 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+WHISPER_OBJ += ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o
+
 whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
@@ -248,14 +324,21 @@ whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-imp
 WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
 endif
 
-libwhisper.a: ggml.o $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+ifdef WHISPER_METAL
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
 
-libwhisper.so: ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
+WHISPER_OBJ += ggml-metal.o
+endif
+
+libwhisper.a: $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a $(WHISPER_OBJ)
+
+libwhisper.so: $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk talk-llama bench quantize libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize server lsp libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -266,27 +349,33 @@ CC_SDL=`sdl2-config --cflags --libs`
 SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
+bench: examples/bench/bench.cpp $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
-	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
+server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
 
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
+
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples
@@ -301,12 +390,19 @@ samples:
 	@wget --quiet --show-progress -O samples/gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
 	@wget --quiet --show-progress -O samples/hp0.ogg https://upload.wikimedia.org/wikipedia/en/d/d4/En.henryfphillips.ogg
 	@wget --quiet --show-progress -O samples/mm1.wav https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav
+	@wget --quiet --show-progress -O samples/a13.mp3 https://upload.wikimedia.org/wikipedia/commons/transcoded/6/6f/Apollo13-wehaveaproblem.ogg/Apollo13-wehaveaproblem.ogg.mp3
+	@wget --quiet --show-progress -O samples/diffusion2023-07-03.flac https://archive.org/download/diffusion2023-07-03/diffusion2023-07-03.flac
 	@echo "Converting to 16-bit WAV ..."
 	@ffmpeg -loglevel -0 -y -i samples/gb0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb0.wav
 	@ffmpeg -loglevel -0 -y -i samples/gb1.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb1.wav
 	@ffmpeg -loglevel -0 -y -i samples/hp0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/hp0.wav
+	@rm samples/*.ogg
 	@ffmpeg -loglevel -0 -y -i samples/mm1.wav -ar 16000 -ac 1 -c:a pcm_s16le samples/mm0.wav
 	@rm samples/mm1.wav
+	@ffmpeg -loglevel -0 -y -i samples/a13.mp3 -ar 16000 -ac 1 -c:a pcm_s16le -ss 00:00:00 -to 00:00:30 samples/a13.wav
+	@rm samples/a13.mp3
+	@ffmpeg -loglevel -0 -y -i samples/diffusion2023-07-03.flac -ar 16000 -ac 1 -c:a pcm_s16le samples/diffusion2023-07-03.wav
+	@rm samples/diffusion2023-07-03.flac
 
 #
 # Models
@@ -324,9 +420,10 @@ samples:
 .PHONY: medium.en
 .PHONY: medium
 .PHONY: large-v1
-.PHONY: large
+.PHONY: large-v2
+.PHONY: large-v3
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="
@@ -348,4 +445,4 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
 
 .PHONY: tests
 tests:
-	bash ./tests/run-tests.sh
+	bash ./tests/run-tests.sh $(word 2, $(MAKECMDGOALS))

Package.swift

@@ -0,0 +1,77 @@
+// swift-tools-version:5.5
+
+import PackageDescription
+
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v12),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let resources: [Resource] = [
+    .process("ggml-metal.metal")
+]
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let resources: [Resource] = []
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
+let package = Package(
+    name: "whisper",
+    platforms: platforms,
+    products: [
+        .library(name: "whisper", targets: ["whisper"]),
+    ],
+    targets: [
+        .target(
+            name: "whisper",
+            path: ".",
+            exclude: exclude + [
+               "bindings",
+               "cmake",
+               "coreml",
+               "examples",
+               "extra",
+               "models",
+               "samples",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
+            sources: [
+                "ggml.c",
+                "whisper.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c"
+            ] + additionalSources,
+            resources: resources,
+            publicHeadersPath: "spm-headers",
+            cSettings: [
+                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+                .define("GGML_USE_ACCELERATE")
+                // NOTE: NEW_LAPACK will required iOS version 16.4+
+                // We should consider add this in the future when we drop support for iOS 14
+                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+                // .define("ACCELERATE_NEW_LAPACK"),
+                // .define("ACCELERATE_LAPACK_ILP64")
+            ] + additionalSettings,
+            linkerSettings: [
+                .linkedFramework("Accelerate")
+            ]
+        )
+    ],
+    cxxLanguageStandard: .cxx11
+)

README.md

@@ -6,21 +6,21 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Beta: [v1.4.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.2) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
-- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
-- Runs on the CPU
-- [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- Support for CPU-only inference
+- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
 - [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
+- [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
 Supported platforms:
@@ -28,15 +28,14 @@ Supported platforms:
 - [x] Mac OS (Intel and Arm)
 - [x] [iOS](examples/whisper.objc)
 - [x] [Android](examples/whisper.android)
+- [x] [Java](bindings/java/README.md)
 - [x] Linux / [FreeBSD](https://github.com/ggerganov/whisper.cpp/issues/56#issuecomment-1350920264)
 - [x] [WebAssembly](examples/whisper.wasm)
 - [x] Windows ([MSVC](https://github.com/ggerganov/whisper.cpp/blob/master/.github/workflows/build.yml#L117-L144) and [MinGW](https://github.com/ggerganov/whisper.cpp/issues/168)]
 - [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/discussions/166)
 
-The entire implementation of the model is contained in 2 source files:
-
-- Tensor operations: [ggml.h](ggml.h) / [ggml.c](ggml.c)
-- Transformer inference: [whisper.h](whisper.h) / [whisper.cpp](whisper.cpp)
+The entire high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
+The rest of the code is part of the [ggml](https://github.com/ggerganov/ggml) machine learning library.
 
 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
 As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device: [whisper.objc](examples/whisper.objc)
@@ -47,6 +46,10 @@ You can also easily make your own offline voice assistant application: [command]
 
 https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
 
+On Apple Silicon, the inference runs fully on the GPU via Metal:
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/c82e8f86-60dc-49f2-b048-d2fdbd6b5225
+
 Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 
 ## Implementation details
@@ -58,7 +61,7 @@ Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 - Various other examples are available in the [examples](examples) folder
 
 The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD
-instrisics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
+intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
 the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
 
 ## Quick start
@@ -106,29 +109,37 @@ options:
   -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
   -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
   -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -sow,      --split-on-word     [false  ] split on word rather than on token
+  -bo N,     --best-of N         [2      ] number of best candidates to keep
   -bs N,     --beam-size N       [-1     ] beam size for beam search
   -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
   -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
   -tr,       --translate         [false  ] translate from source language to english
   -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
   -otxt,     --output-txt        [false  ] output result in a text file
   -ovtt,     --output-vtt        [false  ] output result in a vtt file
   -osrt,     --output-srt        [false  ] output result in a srt file
+  -olrc,     --output-lrc        [false  ] output result in a lrc file
   -owts,     --output-words      [false  ] output script for generating karaoke video
+  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
   -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -nt,       --no-timestamps     [false  ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
              --prompt PROMPT     [       ] initial prompt
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  -ls,       --log-score         [false  ] log best decoder scores of token
 
 
 bash ./models/download-ggml-model.sh base.en
@@ -219,18 +230,19 @@ make small
 make medium.en
 make medium
 make large-v1
-make large
+make large-v2
+make large-v3
 ```
 
 ## Memory usage
 
-| Model  | Disk   | Mem     | SHA                                        |
-| ---    | ---    | ---     | ---                                        |
-| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| Model  | Disk    | Mem      |
+| ---    | ---     | ---      |
+| tiny   |  75 MiB | ~273 MB |
+| base   | 142 MiB | ~388 MB |
+| small  | 466 MiB | ~852 MB |
+| medium | 1.5 GiB | ~2.1 GB |
+| large  | 2.9 GiB | ~3.9 GB |
 
 ## Quantization
 
@@ -283,8 +295,8 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
   WHISPER_COREML=1 make -j
 
   # using CMake
-  cd build
-  cmake -DWHISPER_COREML=1 ..
+  cmake -B build -DWHISPER_COREML=1
+  cmake --build build -j --config Release
   ```
 
 - Run the examples as usual. For example:
@@ -308,9 +320,88 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 
 For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
 
-## NVIDIA GPU support via cuBLAS
+## OpenVINO support
 
-With NVIDIA cards, the Encoder processing can be offloaded to the GPU to a large extend through cuBLAS.
+On platforms that support [OpenVINO](https://github.com/openvinotoolkit/openvino), the Encoder inference can be executed
+on OpenVINO-supported devices including x86 CPUs and Intel GPUs (integrated & discrete).
+
+This can result in significant speedup in encoder performance. Here are the instructions for generating the OpenVINO model and using it with `whisper.cpp`:
+
+- First, setup python virtual env. and install python dependencies. Python 3.10 is recommended.
+
+  Windows:
+  ```
+  cd models
+  python -m venv openvino_conv_env
+  openvino_conv_env\Scripts\activate
+  python -m pip install --upgrade pip
+  pip install -r openvino-conversion-requirements.txt
+  ```
+
+  Linux and macOS:
+  ```
+  cd models
+  python3 -m venv openvino_conv_env
+  source openvino_conv_env/bin/activate
+  python -m pip install --upgrade pip
+  pip install -r openvino-conversion-requirements.txt
+  ```
+
+- Generate an OpenVINO encoder model. For example, to generate a `base.en` model, use:
+
+  ```
+  python convert-whisper-to-openvino.py --model base.en
+  ```
+
+  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as ggml models, as that
+  is the default location that the OpenVINO extension will search at runtime.
+
+- Build `whisper.cpp` with OpenVINO support:
+
+  Download OpenVINO package from [release page](https://github.com/openvinotoolkit/openvino/releases). The recommended version to use is [2023.0.0](https://github.com/openvinotoolkit/openvino/releases/tag/2023.0.0).
+
+  After downloading & extracting package onto your development system, set up required environment by sourcing setupvars script. For example:
+
+  Linux:
+  ```bash
+  source /path/to/l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64/setupvars.sh
+  ```
+
+  Windows (cmd):
+  ```
+  C:\Path\To\w_openvino_toolkit_windows_2023.0.0.10926.b4452d56304_x86_64\setupvars.bat
+  ```
+
+  And then build the project using cmake:
+  ```bash
+  cmake -B build -DWHISPER_OPENVINO=1
+  cmake --build build -j --config Release
+  ```
+
+- Run the examples as usual. For example:
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+
+  ...
+
+  whisper_ctx_init_openvino_encoder: loading OpenVINO model from 'models/ggml-base.en-encoder-openvino.xml'
+  whisper_ctx_init_openvino_encoder: first run on a device may take a while ...
+  whisper_openvino_init: path_model = models/ggml-base.en-encoder-openvino.xml, device = GPU, cache_dir = models/ggml-base.en-encoder-openvino-cache
+  whisper_ctx_init_openvino_encoder: OpenVINO model loaded
+
+  system_info: n_threads = 4 / 8 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | COREML = 0 | OPENVINO = 1 |
+
+  ...
+  ```
+
+  The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
+  cached for the next run.
+
+For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
+
+## NVIDIA GPU support
+
+With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads
 
 Now build `whisper.cpp` with cuBLAS support:
@@ -322,7 +413,7 @@ WHISPER_CUBLAS=1 make -j
 
 ## OpenCL GPU support via CLBlast
 
-For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APU's or low end devices for up to ~2x speedup.
+For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APUs or low end devices for up to ~2x speedup.
 
 First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
 
@@ -335,16 +426,26 @@ make clean
 WHISPER_CLBLAST=1 make -j
 
 CMake:
-cd whisper.cpp ; mkdir build ; cd build
-cmake -DWHISPER_CLBLAST=ON  ..
-make clean
-make -j
-cp bin/* ../ 
+cd whisper.cpp
+cmake -B build -DWHISPER_CLBLAST=ON
+cmake --build build -j --config Release
 ```
 
 
 Run all the examples as usual.
 
+## BLAS CPU support via OpenBLAS
+
+Encoder processing can be accelerated on the CPU via OpenBLAS.
+First, make sure you have installed `openblas`: https://www.openblas.net/
+
+Now build `whisper.cpp` with OpenBLAS support:
+
+```
+make clean
+WHISPER_OPENBLAS=1 make -j
+```
+
 ## Limitations
 
 - Inference only
@@ -479,7 +580,7 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr
 [00:00:10.020 --> 00:00:11.000]   country.
 ```
 
-## Word-level timestamp
+## Word-level timestamp (experimental)
 
 The `--max-len` argument can be used to obtain word-level timestamps. Simply use `-ml 1`:
 
@@ -520,6 +621,32 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr
 [00:00:10.510 --> 00:00:11.000]  .
 ```
 
+## Speaker segmentation via tinydiarize (experimental)
+
+More information about this approach is available here: https://github.com/ggerganov/whisper.cpp/pull/1058
+
+Sample usage:
+
+```py
+# download a tinydiarize compatible model
+./models/download-ggml-model.sh small.en-tdrz
+
+# run as usual, adding the "-tdrz" command-line argument
+./main -f ./samples/a13.wav -m ./models/ggml-small.en-tdrz.bin -tdrz
+...
+main: processing './samples/a13.wav' (480000 samples, 30.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, tdrz = 1, timestamps = 1 ...
+...
+[00:00:00.000 --> 00:00:03.800]   Okay Houston, we've had a problem here. [SPEAKER_TURN]
+[00:00:03.800 --> 00:00:06.200]   This is Houston. Say again please. [SPEAKER_TURN]
+[00:00:06.200 --> 00:00:08.260]   Uh Houston we've had a problem.
+[00:00:08.260 --> 00:00:11.320]   We've had a main beam up on a volt. [SPEAKER_TURN]
+[00:00:11.320 --> 00:00:13.820]   Roger main beam interval. [SPEAKER_TURN]
+[00:00:13.820 --> 00:00:15.100]   Uh uh [SPEAKER_TURN]
+[00:00:15.100 --> 00:00:18.020]   So okay stand, by thirteen we're looking at it. [SPEAKER_TURN]
+[00:00:18.020 --> 00:00:25.740]   Okay uh right now uh Houston the uh voltage is uh is looking good um.
+[00:00:27.620 --> 00:00:29.940]   And we had a a pretty large bank or so.
+```
+
 ## Karaoke-style movie generation (experimental)
 
 The [main](examples/main) example provides support for output of karaoke-style movies, where the
@@ -579,6 +706,19 @@ took to execute it. The results are summarized in the following Github issue:
 
 [Benchmark results](https://github.com/ggerganov/whisper.cpp/issues/89)
 
+Additionally a script to run whisper.cpp with different models and audio files is provided [bench.py](bench.py).
+
+You can run it with the following command, by default it will run against any standard model in the models folder.
+
+```bash
+python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+```
+
+It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
+
+It outputs a csv file with the results of the benchmarking.
+
+
 ## ggml format
 
 The original models are converted to a custom binary format. This allows to pack everything needed into a single file:
@@ -600,9 +740,11 @@ in [models](models).
 ## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
 - [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
-- [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [X] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
   - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
 - [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [X] Java:
+  - [GiviMAD/whisper-jni](https://github.com/GiviMAD/whisper-jni)
 - [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
 - [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
   - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)

bindings/go/Makefile

@@ -32,7 +32,7 @@ mkdir:
 modtidy:
 	@go mod tidy
 
-clean: 
+clean:
 	@echo Clean
 	@rm -fr $(BUILD_DIR)
 	@go clean

bindings/go/README.md

@@ -31,7 +31,7 @@ func main() {
 	if err != nil {
 		panic(err)
 	}
-	if err := context.Process(samples, nil); err != nil {
+	if err := context.Process(samples, nil, nil); err != nil {
 		return err
 	}
 
@@ -71,7 +71,7 @@ The examples are placed in the `build` directory. Once built, you can download a
 And you can then test a model against samples with the following command:
 
 ```bash
-./build/go-whisper -model models/ggml-tiny.en.bin samples/jfk.wav 
+./build/go-whisper -model models/ggml-tiny.en.bin samples/jfk.wav
 ```
 
 ## Using the bindings

bindings/go/examples/go-model-download/main.go

@@ -24,7 +24,7 @@ const (
 
 var (
 	// The models which will be downloaded, if no model is specified as an argument
-	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large"}
+	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large-v3"}
 )
 
 var (

bindings/go/examples/go-whisper/process.go

@@ -67,7 +67,7 @@ func Process(model whisper.Model, path string, flags *Flags) error {
 	// Process the data
 	fmt.Fprintf(flags.Output(), "  ...processing %q\n", path)
 	context.ResetTimings()
-	if err := context.Process(data, cb); err != nil {
+	if err := context.Process(data, cb, nil); err != nil {
 		return err
 	}
 

bindings/go/params.go

@@ -19,6 +19,10 @@ func (p *Params) SetTranslate(v bool) {
 	p.translate = toBool(v)
 }
 
+func (p *Params) SetSplitOnWord(v bool) {
+	p.split_on_word = toBool(v)
+}
+
 func (p *Params) SetNoContext(v bool) {
 	p.no_context = toBool(v)
 }
@@ -114,6 +118,11 @@ func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)
 }
 
+// Set audio encoder context
+func (p *Params) SetAudioCtx(n int) {
+	p.audio_ctx = C.int(n)
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS
 
@@ -137,6 +146,7 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" n_max_text_ctx=%d", p.n_max_text_ctx)
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
+	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
 	if p.translate {
 		str += " translate"
 	}

bindings/go/pkg/whisper/context.go

@@ -81,6 +81,10 @@ func (context *context) SetSpeedup(v bool) {
 	context.params.SetSpeedup(v)
 }
 
+func (context *context) SetSplitOnWord(v bool) {
+	context.params.SetSplitOnWord(v)
+}
+
 // Set number of threads to use
 func (context *context) SetThreads(v uint) {
 	context.params.SetThreads(int(v))
@@ -93,7 +97,7 @@ func (context *context) SetOffset(v time.Duration) {
 
 // Set duration of audio to process
 func (context *context) SetDuration(v time.Duration) {
-	context.params.SetOffset(int(v.Milliseconds()))
+	context.params.SetDuration(int(v.Milliseconds()))
 }
 
 // Set timestamp token probability threshold (~0.01)
@@ -121,6 +125,11 @@ func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
 }
 
+// Set audio encoder context
+func (context *context) SetAudioCtx(n uint) {
+	context.params.SetAudioCtx(int(n))
+}
+
 // ResetTimings resets the mode timings. Should be called before processing
 func (context *context) ResetTimings() {
 	context.model.ctx.Whisper_reset_timings()
@@ -152,12 +161,16 @@ func (context *context) WhisperLangAutoDetect(offset_ms int, n_threads int) ([]f
 }
 
 // Process new sample data and return any errors
-func (context *context) Process(data []float32, cb SegmentCallback) error {
+func (context *context) Process(
+	data []float32,
+	callNewSegment SegmentCallback,
+	callProgress ProgressCallback,
+) error {
 	if context.model.ctx == nil {
 		return ErrInternalAppError
 	}
 	// If the callback is defined then we force on single_segment mode
-	if cb != nil {
+	if callNewSegment != nil {
 		context.params.SetSingleSegment(true)
 	}
 
@@ -165,24 +178,28 @@ func (context *context) Process(data []float32, cb SegmentCallback) error {
 	processors := 0
 	if processors > 1 {
 		if err := context.model.ctx.Whisper_full_parallel(context.params, data, processors, nil, func(new int) {
-			if cb != nil {
+			if callNewSegment != nil {
 				num_segments := context.model.ctx.Whisper_full_n_segments()
 				s0 := num_segments - new
 				for i := s0; i < num_segments; i++ {
-					cb(toSegment(context.model.ctx, i))
+					callNewSegment(toSegment(context.model.ctx, i))
 				}
 			}
 		}); err != nil {
 			return err
 		}
 	} else if err := context.model.ctx.Whisper_full(context.params, data, nil, func(new int) {
-		if cb != nil {
+		if callNewSegment != nil {
 			num_segments := context.model.ctx.Whisper_full_n_segments()
 			s0 := num_segments - new
 			for i := s0; i < num_segments; i++ {
-				cb(toSegment(context.model.ctx, i))
+				callNewSegment(toSegment(context.model.ctx, i))
 			}
 		}
+	}, func(progress int) {
+		if callProgress != nil {
+			callProgress(progress)
+		}
 	}); err != nil {
 		return err
 	}

bindings/go/pkg/whisper/interface.go

@@ -12,6 +12,10 @@ import (
 // time. It is called during the Process function
 type SegmentCallback func(Segment)
 
+// ProgressCallback is the callback function for reporting progress during
+// processing. It is called during the Process function
+type ProgressCallback func(int)
+
 // Model is the interface to a whisper model. Create a new model with the
 // function whisper.New(string)
 type Model interface {
@@ -38,16 +42,18 @@ type Context interface {
 	SetDuration(time.Duration)    // Set duration
 	SetThreads(uint)              // Set number of threads to use
 	SetSpeedup(bool)              // Set speedup flag
+	SetSplitOnWord(bool)          // Set split on word flag
 	SetTokenThreshold(float32)    // Set timestamp token probability threshold
 	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
 	SetMaxSegmentLength(uint)     // Set max segment length in characters
 	SetTokenTimestamps(bool)      // Set token timestamps flag
 	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)
+	SetAudioCtx(uint)             // Set audio encoder context
 
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the
 	// callback function during processing.
-	Process([]float32, SegmentCallback) error
+	Process([]float32, SegmentCallback, ProgressCallback) error
 
 	// After process is called, return segments until the end of the stream
 	// is reached, when io.EOF is returned.

bindings/go/whisper.go

@@ -15,6 +15,7 @@ import (
 #include <stdlib.h>
 
 extern void callNewSegment(void* user_data, int new);
+extern void callProgress(void* user_data, int progress);
 extern bool callEncoderBegin(void* user_data);
 
 // Text segment callback
@@ -26,6 +27,15 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_s
     }
 }
 
+// Progress callback
+// Called on every newly generated text segment
+// Use the whisper_full_...() functions to obtain the text segments
+static void whisper_progress_cb(struct whisper_context* ctx, struct whisper_state* state, int progress, void* user_data) {
+    if(user_data != NULL && ctx != NULL) {
+        callProgress(user_data, progress);
+    }
+}
+
 // Encoder begin callback
 // If not NULL, called before the encoder starts
 // If it returns false, the computation is aborted
@@ -43,6 +53,8 @@ static struct whisper_full_params whisper_full_default_params_cb(struct whisper_
 	params.new_segment_callback_user_data = (void*)(ctx);
 	params.encoder_begin_callback = whisper_encoder_begin_cb;
 	params.encoder_begin_callback_user_data = (void*)(ctx);
+	params.progress_callback = whisper_progress_cb;
+	params.progress_callback_user_data = (void*)(ctx);
 	return params;
 }
 */
@@ -71,7 +83,6 @@ const (
 	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
 	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
 	NumFFT     = C.WHISPER_N_FFT
-	NumMEL     = C.WHISPER_N_MEL
 	HopLength  = C.WHISPER_HOP_LENGTH
 	ChunkSize  = C.WHISPER_CHUNK_SIZE
 )
@@ -91,7 +102,7 @@ var (
 func Whisper_init(path string) *Context {
 	cPath := C.CString(path)
 	defer C.free(unsafe.Pointer(cPath))
-	if ctx := C.whisper_init_from_file(cPath); ctx != nil {
+	if ctx := C.whisper_init_from_file_with_params(cPath, C.whisper_context_default_params()); ctx != nil {
 		return (*Context)(ctx)
 	} else {
 		return nil
@@ -258,13 +269,13 @@ func (ctx *Context) Whisper_token_lang(lang_id int) Token {
 }
 
 // Task tokens
-func Whisper_token_translate() Token {
-	return Token(C.whisper_token_translate())
+func (ctx *Context) Whisper_token_translate() Token {
+	return Token(C.whisper_token_translate((*C.struct_whisper_context)(ctx)))
 }
 
 // Task tokens
-func Whisper_token_transcribe() Token {
-	return Token(C.whisper_token_transcribe())
+func (ctx *Context) Whisper_token_transcribe() Token {
+	return Token(C.whisper_token_transcribe((*C.struct_whisper_context)(ctx)))
 }
 
 // Performance information
@@ -290,11 +301,19 @@ func (ctx *Context) Whisper_full_default_params(strategy SamplingStrategy) Param
 
 // Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
 // Uses the specified decoding strategy to obtain the text.
-func (ctx *Context) Whisper_full(params Params, samples []float32, encoderBeginCallback func() bool, newSegmentCallback func(int)) error {
+func (ctx *Context) Whisper_full(
+	params Params,
+	samples []float32,
+	encoderBeginCallback func() bool,
+	newSegmentCallback func(int),
+	progressCallback func(int),
+) error {
 	registerEncoderBeginCallback(ctx, encoderBeginCallback)
 	registerNewSegmentCallback(ctx, newSegmentCallback)
+	registerProgressCallback(ctx, progressCallback)
 	defer registerEncoderBeginCallback(ctx, nil)
 	defer registerNewSegmentCallback(ctx, nil)
+	defer registerProgressCallback(ctx, nil)
 	if C.whisper_full((*C.struct_whisper_context)(ctx), (C.struct_whisper_full_params)(params), (*C.float)(&samples[0]), C.int(len(samples))) == 0 {
 		return nil
 	} else {
@@ -318,6 +337,18 @@ func (ctx *Context) Whisper_full_parallel(params Params, samples []float32, proc
 	}
 }
 
+// Return the id of the autodetected language, returns -1 if not found
+// Added to whisper.cpp in
+// https://github.com/ggerganov/whisper.cpp/commit/a1c1583cc7cd8b75222857afc936f0638c5683d6
+//
+// Examples:
+//
+//	"de" -> 2
+//	"german" -> 2
+func (ctx *Context) Whisper_full_lang_id() int {
+	return int(C.whisper_full_lang_id((*C.struct_whisper_context)(ctx)))
+}
+
 // Number of generated text segments.
 // A segment can be a few words, a sentence, or even a paragraph.
 func (ctx *Context) Whisper_full_n_segments() int {
@@ -370,6 +401,7 @@ func (ctx *Context) Whisper_full_get_token_p(segment int, token int) float32 {
 
 var (
 	cbNewSegment   = make(map[unsafe.Pointer]func(int))
+	cbProgress     = make(map[unsafe.Pointer]func(int))
 	cbEncoderBegin = make(map[unsafe.Pointer]func() bool)
 )
 
@@ -381,6 +413,14 @@ func registerNewSegmentCallback(ctx *Context, fn func(int)) {
 	}
 }
 
+func registerProgressCallback(ctx *Context, fn func(int)) {
+	if fn == nil {
+		delete(cbProgress, unsafe.Pointer(ctx))
+	} else {
+		cbProgress[unsafe.Pointer(ctx)] = fn
+	}
+}
+
 func registerEncoderBeginCallback(ctx *Context, fn func() bool) {
 	if fn == nil {
 		delete(cbEncoderBegin, unsafe.Pointer(ctx))
@@ -396,6 +436,13 @@ func callNewSegment(user_data unsafe.Pointer, new C.int) {
 	}
 }
 
+//export callProgress
+func callProgress(user_data unsafe.Pointer, progress C.int) {
+	if fn, ok := cbProgress[user_data]; ok {
+		fn(int(progress))
+	}
+}
+
 //export callEncoderBegin
 func callEncoderBegin(user_data unsafe.Pointer) C.bool {
 	if fn, ok := cbEncoderBegin[user_data]; ok {
@@ -415,3 +462,7 @@ func (t TokenData) T0() int64 {
 func (t TokenData) T1() int64 {
 	return int64(t.t1)
 }
+
+func (t TokenData) Id() Token {
+	return Token(t.id)
+}

bindings/go/whisper_test.go

@@ -52,7 +52,7 @@ func Test_Whisper_001(t *testing.T) {
 	defer ctx.Whisper_free()
 	params := ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY)
 	data := buf.AsFloat32Buffer().Data
-	err = ctx.Whisper_full(params, data, nil, nil)
+	err = ctx.Whisper_full(params, data, nil, nil, nil)
 	assert.NoError(err)
 
 	// Print out tokens

bindings/java/.idea/uiDesigner.xml

@@ -0,0 +1,124 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="Palette2">
+    <group name="Swing">
+      <item class="com.intellij.uiDesigner.HSpacer" tooltip-text="Horizontal Spacer" icon="/com/intellij/uiDesigner/icons/hspacer.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="1" hsize-policy="6" anchor="0" fill="1" />
+      </item>
+      <item class="com.intellij.uiDesigner.VSpacer" tooltip-text="Vertical Spacer" icon="/com/intellij/uiDesigner/icons/vspacer.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="1" anchor="0" fill="2" />
+      </item>
+      <item class="javax.swing.JPanel" icon="/com/intellij/uiDesigner/icons/panel.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="3" hsize-policy="3" anchor="0" fill="3" />
+      </item>
+      <item class="javax.swing.JScrollPane" icon="/com/intellij/uiDesigner/icons/scrollPane.svg" removable="false" auto-create-binding="false" can-attach-label="true">
+        <default-constraints vsize-policy="7" hsize-policy="7" anchor="0" fill="3" />
+      </item>
+      <item class="javax.swing.JButton" icon="/com/intellij/uiDesigner/icons/button.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="3" anchor="0" fill="1" />
+        <initial-values>
+          <property name="text" value="Button" />
+        </initial-values>
+      </item>
+      <item class="javax.swing.JRadioButton" icon="/com/intellij/uiDesigner/icons/radioButton.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="3" anchor="8" fill="0" />
+        <initial-values>
+          <property name="text" value="RadioButton" />
+        </initial-values>
+      </item>
+      <item class="javax.swing.JCheckBox" icon="/com/intellij/uiDesigner/icons/checkBox.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="3" anchor="8" fill="0" />
+        <initial-values>
+          <property name="text" value="CheckBox" />
+        </initial-values>
+      </item>
+      <item class="javax.swing.JLabel" icon="/com/intellij/uiDesigner/icons/label.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="0" anchor="8" fill="0" />
+        <initial-values>
+          <property name="text" value="Label" />
+        </initial-values>
+      </item>
+      <item class="javax.swing.JTextField" icon="/com/intellij/uiDesigner/icons/textField.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="8" fill="1">
+          <preferred-size width="150" height="-1" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JPasswordField" icon="/com/intellij/uiDesigner/icons/passwordField.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="8" fill="1">
+          <preferred-size width="150" height="-1" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JFormattedTextField" icon="/com/intellij/uiDesigner/icons/formattedTextField.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="8" fill="1">
+          <preferred-size width="150" height="-1" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JTextArea" icon="/com/intellij/uiDesigner/icons/textArea.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JTextPane" icon="/com/intellij/uiDesigner/icons/textPane.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JEditorPane" icon="/com/intellij/uiDesigner/icons/editorPane.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JComboBox" icon="/com/intellij/uiDesigner/icons/comboBox.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="0" hsize-policy="2" anchor="8" fill="1" />
+      </item>
+      <item class="javax.swing.JTable" icon="/com/intellij/uiDesigner/icons/table.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JList" icon="/com/intellij/uiDesigner/icons/list.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="2" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JTree" icon="/com/intellij/uiDesigner/icons/tree.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3">
+          <preferred-size width="150" height="50" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JTabbedPane" icon="/com/intellij/uiDesigner/icons/tabbedPane.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="3" hsize-policy="3" anchor="0" fill="3">
+          <preferred-size width="200" height="200" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JSplitPane" icon="/com/intellij/uiDesigner/icons/splitPane.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="3" hsize-policy="3" anchor="0" fill="3">
+          <preferred-size width="200" height="200" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JSpinner" icon="/com/intellij/uiDesigner/icons/spinner.svg" removable="false" auto-create-binding="true" can-attach-label="true">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="8" fill="1" />
+      </item>
+      <item class="javax.swing.JSlider" icon="/com/intellij/uiDesigner/icons/slider.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="8" fill="1" />
+      </item>
+      <item class="javax.swing.JSeparator" icon="/com/intellij/uiDesigner/icons/separator.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="6" anchor="0" fill="3" />
+      </item>
+      <item class="javax.swing.JProgressBar" icon="/com/intellij/uiDesigner/icons/progressbar.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="0" fill="1" />
+      </item>
+      <item class="javax.swing.JToolBar" icon="/com/intellij/uiDesigner/icons/toolbar.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="6" anchor="0" fill="1">
+          <preferred-size width="-1" height="20" />
+        </default-constraints>
+      </item>
+      <item class="javax.swing.JToolBar$Separator" icon="/com/intellij/uiDesigner/icons/toolbarSeparator.svg" removable="false" auto-create-binding="false" can-attach-label="false">
+        <default-constraints vsize-policy="0" hsize-policy="0" anchor="0" fill="1" />
+      </item>
+      <item class="javax.swing.JScrollBar" icon="/com/intellij/uiDesigner/icons/scrollbar.svg" removable="false" auto-create-binding="true" can-attach-label="false">
+        <default-constraints vsize-policy="6" hsize-policy="0" anchor="0" fill="2" />
+      </item>
+    </group>
+  </component>
+</project>

bindings/java/README.md

@@ -0,0 +1,71 @@
+# Java JNI bindings for Whisper
+
+This package provides Java JNI bindings for whisper.cpp. They have been tested on:
+
+  * <strike>Darwin (OS X) 12.6 on x64_64</strike>
+  * Ubuntu on x86_64
+  * Windows on x86_64
+
+The "low level" bindings are in `WhisperCppJnaLibrary`. The most simple usage is as follows:
+
+JNA will attempt to load the `whispercpp` shared library from:
+
+- jna.library.path
+- jna.platform.library
+- ~/Library/Frameworks
+- /Library/Frameworks
+- /System/Library/Frameworks
+- classpath
+
+```java
+import io.github.ggerganov.whispercpp.WhisperCpp;
+
+public class Example {
+
+    public static void main(String[] args) {
+        WhisperCpp whisper = new WhisperCpp();
+        // By default, models are loaded from ~/.cache/whisper/ and are usually named "ggml-${name}.bin"
+        // or you can provide the absolute path to the model file.
+        long context = whisper.initContext("base.en");
+        try {
+            var whisperParams = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+            // custom configuration if required
+            whisperParams.temperature_inc = 0f;
+            
+            var samples = readAudio(); // divide each value by 32767.0f
+            whisper.fullTranscribe(whisperParams, samples);
+            
+            int segmentCount = whisper.getTextSegmentCount(context);
+            for (int i = 0; i < segmentCount; i++) {
+                String text = whisper.getTextSegment(context, i);
+                System.out.println(segment.getText());
+            }
+        } finally {
+             whisper.freeContext(context);
+        }
+     }
+}
+```
+
+## Building & Testing
+
+In order to build, you need to have the JDK 8 or higher installed. Run the tests with:
+
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+cd whisper.cpp/bindings/java
+
+./gradlew build
+```
+
+You need to have the `whisper` library in your [JNA library path](https://java-native-access.github.io/jna/4.2.1/com/sun/jna/NativeLibrary.html). On Windows the dll is included in the jar and you can update it:
+
+```bash
+copy /y ..\..\build\bin\Release\whisper.dll build\generated\resources\main\win32-x86-64\whisper.dll
+```
+
+
+## License
+
+The license for the Go bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
+

bindings/java/build.gradle

@@ -0,0 +1,133 @@
+plugins {
+    id 'java'
+    id 'java-library'
+    id 'maven-publish'
+    id 'signing'
+}
+
+archivesBaseName = 'whispercpp'
+group = 'io.github.ggerganov'
+version = '1.4.0'
+
+
+sourceCompatibility = 1.8
+targetCompatibility = 1.8
+
+sourceSets {
+    main {
+        resources {
+            srcDirs = ['src/main/resources', 'build/generated/resources/main']
+        }
+    }
+    test {
+        runtimeClasspath += files('build/generated/resources/main')
+    }
+}
+
+tasks.register('copyLibwhisperDynlib', Copy) {
+    from '../../build'
+    include 'libwhisper.dynlib'
+    into 'build/generated/resources/main/darwin'
+}
+
+tasks.register('copyLibwhisperSo', Copy) {
+    from '../../build'
+    include 'libwhisper.so'
+    into 'build/generated/resources/main/linux-x86-64'
+}
+
+tasks.register('copyWhisperDll', Copy) {
+    from '../../build/Release'
+    include 'whisper.dll'
+    into 'build/generated/resources/main/windows-x86-64'
+}
+
+tasks.register('copyLibs') {
+    dependsOn copyLibwhisperDynlib, copyLibwhisperSo, copyWhisperDll
+}
+
+test {
+    systemProperty 'jna.library.path', project.file('build/generated/resources/main').absolutePath
+}
+
+java {
+    withSourcesJar()
+    withJavadocJar()
+}
+
+jar {
+    exclude '**/whisper_java.exp', '**/whisper_java.lib'
+}
+
+javadoc {
+    options.addStringOption('Xdoclint:none', '-quiet')
+}
+
+tasks.withType(Test) {
+    useJUnitPlatform()
+}
+
+dependencies {
+    implementation "net.java.dev.jna:jna:5.13.0"
+    testImplementation "org.junit.jupiter:junit-jupiter:5.9.2"
+    testImplementation "org.assertj:assertj-core:3.24.2"
+}
+
+repositories {
+    mavenCentral()
+}
+
+publishing {
+    publications {
+        mavenJava(MavenPublication) {
+            artifactId = 'whispercpp'
+            from components.java
+            pom {
+                name = 'whispercpp'
+                description = "Java JNA bindings for OpenAI's Whisper model, implemented in C/C++"
+                url = 'https://github.com/ggerganov/whisper.cpp'
+                licenses {
+                    license {
+                        name = 'MIT licence'
+                        url = 'https://raw.githubusercontent.com/ggerganov/whisper.cpp/master/LICENSE'
+                    }
+                }
+                developers {
+                    developer {
+                        id = 'ggerganov'
+                        name = 'Georgi Gerganov'
+                        email = 'ggerganov@gmail.com'
+                    }
+                    developer {
+                        id = 'nalbion'
+                        name = 'Nicholas Albion'
+                        email = 'nalbion@yahoo.com'
+                    }
+                }
+                scm {
+                    connection = 'scm:git:git://github.com/ggerganov/whisper.cpp.git'
+                    url = 'https://github.com/ggerganov/whisper.cpp'
+                }
+            }
+        }
+    }
+
+    repositories {
+        maven {
+            def releasesRepoUrl = 'https://s01.oss.sonatype.org/service/local/staging/deploy/maven2/'
+            def snapshotsRepoUrl = 'https://s01.oss.sonatype.org/content/repositories/snapshots/'
+            url = version.endsWith('-SNAPSHOT') ? snapshotsRepoUrl : releasesRepoUrl
+            credentials {
+                username = System.getenv("MAVEN_USERNAME")
+                password = System.getenv("MAVEN_PASSWORD")
+            }
+        }
+    }
+}
+
+signing {
+    def signingKey = System.getenv("PGP_SECRET")
+    def signingPassword = System.getenv("PGP_PASSPHRASE")
+    useInMemoryPgpKeys(signingKey, signingPassword)
+    sign publishing.publications.mavenJava
+}

bindings/java/gradle.properties

@@ -0,0 +1,6 @@
+org.gradle.jvmargs=-Xms256m -Xmx1024m
+system.include.dir=/usr/include
+#system.local.include.dir=../../include
+system.local.include.dir=./build/generated/sources/headers/java/main
+jni.include.dir=/usr/lib/jvm/java-8-openjdk-amd64/include/
+jni.lib.dir=/usr/lib/jvm/java-8-openjdk-amd64/lib/

bindings/java/gradle/wrapper/gradle-wrapper.properties

@@ -0,0 +1,6 @@
+distributionBase=GRADLE_USER_HOME
+distributionPath=wrapper/dists
+distributionUrl=https\://services.gradle.org/distributions/gradle-8.1-bin.zip
+networkTimeout=10000
+zipStoreBase=GRADLE_USER_HOME
+zipStorePath=wrapper/dists

bindings/java/gradlew

@@ -0,0 +1,244 @@
+#!/bin/sh
+
+#
+# Copyright © 2015-2021 the original authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+##############################################################################
+#
+#   Gradle start up script for POSIX generated by Gradle.
+#
+#   Important for running:
+#
+#   (1) You need a POSIX-compliant shell to run this script. If your /bin/sh is
+#       noncompliant, but you have some other compliant shell such as ksh or
+#       bash, then to run this script, type that shell name before the whole
+#       command line, like:
+#
+#           ksh Gradle
+#
+#       Busybox and similar reduced shells will NOT work, because this script
+#       requires all of these POSIX shell features:
+#         * functions;
+#         * expansions «$var», «${var}», «${var:-default}», «${var+SET}»,
+#           «${var#prefix}», «${var%suffix}», and «$( cmd )»;
+#         * compound commands having a testable exit status, especially «case»;
+#         * various built-in commands including «command», «set», and «ulimit».
+#
+#   Important for patching:
+#
+#   (2) This script targets any POSIX shell, so it avoids extensions provided
+#       by Bash, Ksh, etc; in particular arrays are avoided.
+#
+#       The "traditional" practice of packing multiple parameters into a
+#       space-separated string is a well documented source of bugs and security
+#       problems, so this is (mostly) avoided, by progressively accumulating
+#       options in "$@", and eventually passing that to Java.
+#
+#       Where the inherited environment variables (DEFAULT_JVM_OPTS, JAVA_OPTS,
+#       and GRADLE_OPTS) rely on word-splitting, this is performed explicitly;
+#       see the in-line comments for details.
+#
+#       There are tweaks for specific operating systems such as AIX, CygWin,
+#       Darwin, MinGW, and NonStop.
+#
+#   (3) This script is generated from the Groovy template
+#       https://github.com/gradle/gradle/blob/HEAD/subprojects/plugins/src/main/resources/org/gradle/api/internal/plugins/unixStartScript.txt
+#       within the Gradle project.
+#
+#       You can find Gradle at https://github.com/gradle/gradle/.
+#
+##############################################################################
+
+# Attempt to set APP_HOME
+
+# Resolve links: $0 may be a link
+app_path=$0
+
+# Need this for daisy-chained symlinks.
+while
+    APP_HOME=${app_path%"${app_path##*/}"}  # leaves a trailing /; empty if no leading path
+    [ -h "$app_path" ]
+do
+    ls=$( ls -ld "$app_path" )
+    link=${ls#*' -> '}
+    case $link in             #(
+      /*)   app_path=$link ;; #(
+      *)    app_path=$APP_HOME$link ;;
+    esac
+done
+
+# This is normally unused
+# shellcheck disable=SC2034
+APP_BASE_NAME=${0##*/}
+APP_HOME=$( cd "${APP_HOME:-./}" && pwd -P ) || exit
+
+# Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+DEFAULT_JVM_OPTS='"-Xmx64m" "-Xms64m"'
+
+# Use the maximum available, or set MAX_FD != -1 to use that value.
+MAX_FD=maximum
+
+warn () {
+    echo "$*"
+} >&2
+
+die () {
+    echo
+    echo "$*"
+    echo
+    exit 1
+} >&2
+
+# OS specific support (must be 'true' or 'false').
+cygwin=false
+msys=false
+darwin=false
+nonstop=false
+case "$( uname )" in                #(
+  CYGWIN* )         cygwin=true  ;; #(
+  Darwin* )         darwin=true  ;; #(
+  MSYS* | MINGW* )  msys=true    ;; #(
+  NONSTOP* )        nonstop=true ;;
+esac
+
+CLASSPATH=$APP_HOME/gradle/wrapper/gradle-wrapper.jar
+
+
+# Determine the Java command to use to start the JVM.
+if [ -n "$JAVA_HOME" ] ; then
+    if [ -x "$JAVA_HOME/jre/sh/java" ] ; then
+        # IBM's JDK on AIX uses strange locations for the executables
+        JAVACMD=$JAVA_HOME/jre/sh/java
+    else
+        JAVACMD=$JAVA_HOME/bin/java
+    fi
+    if [ ! -x "$JAVACMD" ] ; then
+        die "ERROR: JAVA_HOME is set to an invalid directory: $JAVA_HOME
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+    fi
+else
+    JAVACMD=java
+    which java >/dev/null 2>&1 || die "ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+
+Please set the JAVA_HOME variable in your environment to match the
+location of your Java installation."
+fi
+
+# Increase the maximum file descriptors if we can.
+if ! "$cygwin" && ! "$darwin" && ! "$nonstop" ; then
+    case $MAX_FD in #(
+      max*)
+        # In POSIX sh, ulimit -H is undefined. That's why the result is checked to see if it worked.
+        # shellcheck disable=SC3045
+        MAX_FD=$( ulimit -H -n ) ||
+            warn "Could not query maximum file descriptor limit"
+    esac
+    case $MAX_FD in  #(
+      '' | soft) :;; #(
+      *)
+        # In POSIX sh, ulimit -n is undefined. That's why the result is checked to see if it worked.
+        # shellcheck disable=SC3045
+        ulimit -n "$MAX_FD" ||
+            warn "Could not set maximum file descriptor limit to $MAX_FD"
+    esac
+fi
+
+# Collect all arguments for the java command, stacking in reverse order:
+#   * args from the command line
+#   * the main class name
+#   * -classpath
+#   * -D...appname settings
+#   * --module-path (only if needed)
+#   * DEFAULT_JVM_OPTS, JAVA_OPTS, and GRADLE_OPTS environment variables.
+
+# For Cygwin or MSYS, switch paths to Windows format before running java
+if "$cygwin" || "$msys" ; then
+    APP_HOME=$( cygpath --path --mixed "$APP_HOME" )
+    CLASSPATH=$( cygpath --path --mixed "$CLASSPATH" )
+
+    JAVACMD=$( cygpath --unix "$JAVACMD" )
+
+    # Now convert the arguments - kludge to limit ourselves to /bin/sh
+    for arg do
+        if
+            case $arg in                                #(
+              -*)   false ;;                            # don't mess with options #(
+              /?*)  t=${arg#/} t=/${t%%/*}              # looks like a POSIX filepath
+                    [ -e "$t" ] ;;                      #(
+              *)    false ;;
+            esac
+        then
+            arg=$( cygpath --path --ignore --mixed "$arg" )
+        fi
+        # Roll the args list around exactly as many times as the number of
+        # args, so each arg winds up back in the position where it started, but
+        # possibly modified.
+        #
+        # NB: a `for` loop captures its iteration list before it begins, so
+        # changing the positional parameters here affects neither the number of
+        # iterations, nor the values presented in `arg`.
+        shift                   # remove old arg
+        set -- "$@" "$arg"      # push replacement arg
+    done
+fi
+
+# Collect all arguments for the java command;
+#   * $DEFAULT_JVM_OPTS, $JAVA_OPTS, and $GRADLE_OPTS can contain fragments of
+#     shell script including quotes and variable substitutions, so put them in
+#     double quotes to make sure that they get re-expanded; and
+#   * put everything else in single quotes, so that it's not re-expanded.
+
+set -- \
+        "-Dorg.gradle.appname=$APP_BASE_NAME" \
+        -classpath "$CLASSPATH" \
+        org.gradle.wrapper.GradleWrapperMain \
+        "$@"
+
+# Stop when "xargs" is not available.
+if ! command -v xargs >/dev/null 2>&1
+then
+    die "xargs is not available"
+fi
+
+# Use "xargs" to parse quoted args.
+#
+# With -n1 it outputs one arg per line, with the quotes and backslashes removed.
+#
+# In Bash we could simply go:
+#
+#   readarray ARGS < <( xargs -n1 <<<"$var" ) &&
+#   set -- "${ARGS[@]}" "$@"
+#
+# but POSIX shell has neither arrays nor command substitution, so instead we
+# post-process each arg (as a line of input to sed) to backslash-escape any
+# character that might be a shell metacharacter, then use eval to reverse
+# that process (while maintaining the separation between arguments), and wrap
+# the whole thing up as a single "set" statement.
+#
+# This will of course break if any of these variables contains a newline or
+# an unmatched quote.
+#
+
+eval "set -- $(
+        printf '%s\n' "$DEFAULT_JVM_OPTS $JAVA_OPTS $GRADLE_OPTS" |
+        xargs -n1 |
+        sed ' s~[^-[:alnum:]+,./:=@_]~\\&~g; ' |
+        tr '\n' ' '
+    )" '"$@"'
+
+exec "$JAVACMD" "$@"

bindings/java/gradlew.bat

@@ -0,0 +1,92 @@
+@rem
+@rem Copyright 2015 the original author or authors.
+@rem
+@rem Licensed under the Apache License, Version 2.0 (the "License");
+@rem you may not use this file except in compliance with the License.
+@rem You may obtain a copy of the License at
+@rem
+@rem      https://www.apache.org/licenses/LICENSE-2.0
+@rem
+@rem Unless required by applicable law or agreed to in writing, software
+@rem distributed under the License is distributed on an "AS IS" BASIS,
+@rem WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+@rem See the License for the specific language governing permissions and
+@rem limitations under the License.
+@rem
+
+@if "%DEBUG%"=="" @echo off
+@rem ##########################################################################
+@rem
+@rem  Gradle startup script for Windows
+@rem
+@rem ##########################################################################
+
+@rem Set local scope for the variables with windows NT shell
+if "%OS%"=="Windows_NT" setlocal
+
+set DIRNAME=%~dp0
+if "%DIRNAME%"=="" set DIRNAME=.
+@rem This is normally unused
+set APP_BASE_NAME=%~n0
+set APP_HOME=%DIRNAME%
+
+@rem Resolve any "." and ".." in APP_HOME to make it shorter.
+for %%i in ("%APP_HOME%") do set APP_HOME=%%~fi
+
+@rem Add default JVM options here. You can also use JAVA_OPTS and GRADLE_OPTS to pass JVM options to this script.
+set DEFAULT_JVM_OPTS="-Xmx64m" "-Xms64m"
+
+@rem Find java.exe
+if defined JAVA_HOME goto findJavaFromJavaHome
+
+set JAVA_EXE=java.exe
+%JAVA_EXE% -version >NUL 2>&1
+if %ERRORLEVEL% equ 0 goto execute
+
+echo.
+echo ERROR: JAVA_HOME is not set and no 'java' command could be found in your PATH.
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:findJavaFromJavaHome
+set JAVA_HOME=%JAVA_HOME:"=%
+set JAVA_EXE=%JAVA_HOME%/bin/java.exe
+
+if exist "%JAVA_EXE%" goto execute
+
+echo.
+echo ERROR: JAVA_HOME is set to an invalid directory: %JAVA_HOME%
+echo.
+echo Please set the JAVA_HOME variable in your environment to match the
+echo location of your Java installation.
+
+goto fail
+
+:execute
+@rem Setup the command line
+
+set CLASSPATH=%APP_HOME%\gradle\wrapper\gradle-wrapper.jar
+
+
+@rem Execute Gradle
+"%JAVA_EXE%" %DEFAULT_JVM_OPTS% %JAVA_OPTS% %GRADLE_OPTS% "-Dorg.gradle.appname=%APP_BASE_NAME%" -classpath "%CLASSPATH%" org.gradle.wrapper.GradleWrapperMain %*
+
+:end
+@rem End local scope for the variables with windows NT shell
+if %ERRORLEVEL% equ 0 goto mainEnd
+
+:fail
+rem Set variable GRADLE_EXIT_CONSOLE if you need the _script_ return code instead of
+rem the _cmd.exe /c_ return code!
+set EXIT_CODE=%ERRORLEVEL%
+if %EXIT_CODE% equ 0 set EXIT_CODE=1
+if not ""=="%GRADLE_EXIT_CONSOLE%" exit %EXIT_CODE%
+exit /b %EXIT_CODE%
+
+:mainEnd
+if "%OS%"=="Windows_NT" endlocal
+
+:omega

bindings/java/settings.gradle

@@ -0,0 +1 @@
+rootProject.name = "whispercpp"

bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperContext.java

@@ -0,0 +1,41 @@
+package io.github.ggerganov.whispercpp;
+
+import com.sun.jna.Structure;
+import com.sun.jna.ptr.PointerByReference;
+import io.github.ggerganov.whispercpp.ggml.GgmlType;
+import io.github.ggerganov.whispercpp.WhisperModel;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
+
+import java.util.List;
+
+public class WhisperContext extends Structure {
+    int t_load_us = 0;
+    int t_start_us = 0;
+
+    /** weight type (FP32 / FP16 / QX) */
+    GgmlType wtype = GgmlType.GGML_TYPE_F16;
+    /** intermediate type (FP32 or FP16) */
+    GgmlType itype = GgmlType.GGML_TYPE_F16;
+
+//    WhisperModel model;
+    public PointerByReference model;
+//    whisper_vocab vocab;
+//    whisper_state * state = nullptr;
+    public PointerByReference vocab;
+    public PointerByReference state;
+
+    /** populated by whisper_init_from_file_with_params() */
+    String path_model;
+    WhisperContextParams params;
+
+//    public static class ByReference extends WhisperContext implements Structure.ByReference {
+//    }
+//
+//    public static class ByValue extends WhisperContext implements Structure.ByValue {
+//    }
+//
+//    @Override
+//    protected List<String> getFieldOrder() {
+//        return List.of("t_load_us", "t_start_us", "wtype", "itype", "model", "vocab", "state", "path_model");
+//    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCpp.java

@@ -0,0 +1,207 @@
+package io.github.ggerganov.whispercpp;
+
+import com.sun.jna.Native;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.bean.WhisperSegment;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
+import io.github.ggerganov.whispercpp.params.WhisperFullParams;
+import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
+
+import java.io.File;
+import java.io.FileNotFoundException;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * Before calling most methods, you must call `initContext(modelPath)` to initialise the `ctx` Pointer.
+ */
+public class WhisperCpp implements AutoCloseable {
+    private WhisperCppJnaLibrary lib = WhisperCppJnaLibrary.instance;
+    private Pointer ctx = null;
+    private Pointer paramsPointer = null;
+    private Pointer greedyParamsPointer = null;
+    private Pointer beamParamsPointer = null;
+
+    public File modelDir() {
+        String modelDirPath = System.getenv("XDG_CACHE_HOME");
+        if (modelDirPath == null) {
+            modelDirPath = System.getProperty("user.home") + "/.cache";
+        }
+
+        return new File(modelDirPath, "whisper");
+    }
+
+    /**
+     * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
+     */
+    public void initContext(String modelPath) throws FileNotFoundException {
+        initContextImpl(modelPath, getContextDefaultParams());
+    }
+
+    /**
+     * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
+     * @param params - params to use when initialising the context
+     */
+    public void initContext(String modelPath, WhisperContextParams params) throws FileNotFoundException {
+        initContextImpl(modelPath, params);
+    }
+
+    private void initContextImpl(String modelPath, WhisperContextParams params) throws FileNotFoundException {
+        if (ctx != null) {
+            lib.whisper_free(ctx);
+        }
+
+        if (!modelPath.contains("/") && !modelPath.contains("\\")) {
+            if (!modelPath.endsWith(".bin")) {
+                modelPath = "ggml-" + modelPath.replace("-", ".") + ".bin";
+            }
+
+            modelPath = new File(modelDir(), modelPath).getAbsolutePath();
+        }
+
+        ctx = lib.whisper_init_from_file_with_params(modelPath, params);
+
+        if (ctx == null) {
+            throw new FileNotFoundException(modelPath);
+        }
+    }
+
+    /**
+     * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
+     * Because this function allocates memory for the params, the caller must call either:
+     * - call `whisper_free_context_params()`
+     * - `Native.free(Pointer.nativeValue(pointer));`
+     */
+    public WhisperContextParams getContextDefaultParams() {
+        paramsPointer = lib.whisper_context_default_params_by_ref();
+        WhisperContextParams params = new WhisperContextParams(paramsPointer);
+        params.read();
+        return params;
+    }
+    
+    /**
+     * Provides default params which can be used with `whisper_full()` etc.
+     * Because this function allocates memory for the params, the caller must call either:
+     * - call `whisper_free_params()`
+     * - `Native.free(Pointer.nativeValue(pointer));`
+     *
+     * @param strategy - GREEDY
+     */
+    public WhisperFullParams getFullDefaultParams(WhisperSamplingStrategy strategy) {
+        Pointer pointer;
+
+        // whisper_full_default_params_by_ref allocates memory which we need to delete, so only create max 1 pointer for each strategy.
+        if (strategy == WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY) {
+            if (greedyParamsPointer == null) {
+                greedyParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            }
+            pointer = greedyParamsPointer;
+        } else {
+            if (beamParamsPointer == null) {
+                beamParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            }
+            pointer = beamParamsPointer;
+        }
+
+        WhisperFullParams params = new WhisperFullParams(pointer);
+        params.read();
+        return params;
+    }
+
+    @Override
+    public void close() {
+        freeContext();
+        freeParams();
+        System.out.println("Whisper closed");
+    }
+
+    private void freeContext() {
+        if (ctx != null) {
+            lib.whisper_free(ctx);
+        }
+    }
+
+    private void freeParams() {
+        if (paramsPointer != null) {
+            Native.free(Pointer.nativeValue(paramsPointer));
+            paramsPointer = null;
+        }
+        if (greedyParamsPointer != null) {
+            Native.free(Pointer.nativeValue(greedyParamsPointer));
+            greedyParamsPointer = null;
+        }
+        if (beamParamsPointer != null) {
+            Native.free(Pointer.nativeValue(beamParamsPointer));
+            beamParamsPointer = null;
+        }
+    }
+
+    /**
+     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text.
+     * Not thread safe for same context
+     * Uses the specified decoding strategy to obtain the text.
+     */
+    public String fullTranscribe(WhisperFullParams whisperParams, float[] audioData) throws IOException {
+        if (ctx == null) {
+            throw new IllegalStateException("Model not initialised");
+        }
+
+        if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
+            throw new IOException("Failed to process audio");
+        }
+
+        int nSegments = lib.whisper_full_n_segments(ctx);
+
+        StringBuilder str = new StringBuilder();
+
+        for (int i = 0; i < nSegments; i++) {
+            String text = lib.whisper_full_get_segment_text(ctx, i);
+            System.out.println("Segment:" + text);
+            str.append(text);
+        }
+
+        return str.toString().trim();
+    }
+    public List<WhisperSegment> fullTranscribeWithTime(WhisperFullParams whisperParams, float[] audioData) throws IOException {
+        if (ctx == null) {
+            throw new IllegalStateException("Model not initialised");
+        }
+
+        if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
+            throw new IOException("Failed to process audio");
+        }
+
+        int nSegments = lib.whisper_full_n_segments(ctx);
+        List<WhisperSegment> segments= new ArrayList<>(nSegments);
+
+
+        for (int i = 0; i < nSegments; i++) {
+            long t0 = lib.whisper_full_get_segment_t0(ctx, i);
+            String text = lib.whisper_full_get_segment_text(ctx, i);
+            long t1 = lib.whisper_full_get_segment_t1(ctx, i);
+            segments.add(new WhisperSegment(t0,t1,text));
+        }
+
+        return segments;
+    }
+
+//    public int getTextSegmentCount(Pointer ctx) {
+//        return lib.whisper_full_n_segments(ctx);
+//    }
+//    public String getTextSegment(Pointer ctx, int index) {
+//        return lib.whisper_full_get_segment_text(ctx, index);
+//    }
+
+    public String getSystemInfo() {
+        return lib.whisper_print_system_info();
+    }
+
+    public int benchMemcpy(int nthread) {
+        return lib.whisper_bench_memcpy(nthread);
+    }
+
+    public int benchGgmlMulMat(int nthread) {
+        return lib.whisper_bench_ggml_mul_mat(nthread);
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java

@@ -0,0 +1,396 @@
+package io.github.ggerganov.whispercpp;
+
+import com.sun.jna.Library;
+import com.sun.jna.Native;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.model.WhisperModelLoader;
+import io.github.ggerganov.whispercpp.model.WhisperTokenData;
+import io.github.ggerganov.whispercpp.params.WhisperContextParams;
+import io.github.ggerganov.whispercpp.params.WhisperFullParams;
+
+public interface WhisperCppJnaLibrary extends Library {
+    WhisperCppJnaLibrary instance = Native.load("whisper", WhisperCppJnaLibrary.class);
+
+    String whisper_print_system_info();
+
+    /**
+     * DEPRECATED. Allocate (almost) all memory needed for the model by loading from a file.
+     *
+     * @param path_model Path to the model file
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_from_file(String path_model);
+    
+    /**
+     * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
+     * Because this function allocates memory for the params, the caller must call either:
+     * - call `whisper_free_context_params()`
+     * - `Native.free(Pointer.nativeValue(pointer));`
+     */
+    Pointer whisper_context_default_params_by_ref();
+
+    void whisper_free_context_params(Pointer params);
+
+    /**
+     * Allocate (almost) all memory needed for the model by loading from a file.
+     *
+     * @param path_model Path to the model file
+     * @param params     Pointer to whisper_context_params
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_from_file_with_params(String path_model, WhisperContextParams params);
+
+    /**
+     * Allocate (almost) all memory needed for the model by loading from a buffer.
+     *
+     * @param buffer       Model buffer
+     * @param buffer_size  Size of the model buffer
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_from_buffer(Pointer buffer, int buffer_size);
+
+    /**
+     * Allocate (almost) all memory needed for the model using a model loader.
+     *
+     * @param loader Model loader
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init(WhisperModelLoader loader);
+
+    /**
+     * Allocate (almost) all memory needed for the model by loading from a file without allocating the state.
+     *
+     * @param path_model Path to the model file
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_from_file_no_state(String path_model);
+
+    /**
+     * Allocate (almost) all memory needed for the model by loading from a buffer without allocating the state.
+     *
+     * @param buffer       Model buffer
+     * @param buffer_size  Size of the model buffer
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_from_buffer_no_state(Pointer buffer, int buffer_size);
+
+//    Pointer whisper_init_from_buffer_no_state(Pointer buffer, long buffer_size);
+
+    /**
+     * Allocate (almost) all memory needed for the model using a model loader without allocating the state.
+     *
+     * @param loader Model loader
+     * @return Whisper context on success, null on failure
+     */
+    Pointer whisper_init_no_state(WhisperModelLoader loader);
+
+    /**
+     * Allocate memory for the Whisper state.
+     *
+     * @param ctx Whisper context
+     * @return Whisper state on success, null on failure
+     */
+    Pointer whisper_init_state(Pointer ctx);
+
+    /**
+     * Free all allocated memory associated with the Whisper context.
+     *
+     * @param ctx Whisper context
+     */
+    void whisper_free(Pointer ctx);
+
+    /**
+     * Free all allocated memory associated with the Whisper state.
+     *
+     * @param state Whisper state
+     */
+    void whisper_free_state(Pointer state);
+
+
+    /**
+     * Convert RAW PCM audio to log mel spectrogram.
+     * The resulting spectrogram is stored inside the default state of the provided whisper context.
+     *
+     * @param ctx - Pointer to a WhisperContext
+     * @return 0 on success
+     */
+    int whisper_pcm_to_mel(Pointer ctx, final float[] samples, int n_samples, int n_threads);
+
+    /**
+     * @param ctx Pointer to a WhisperContext
+     * @param state Pointer to WhisperState
+     * @param n_samples
+     * @param n_threads
+     * @return 0 on success
+     */
+    int whisper_pcm_to_mel_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);
+
+    /**
+     * This can be used to set a custom log mel spectrogram inside the default state of the provided whisper context.
+     * Use this instead of whisper_pcm_to_mel() if you want to provide your own log mel spectrogram.
+     * n_mel must be 80
+     * @return 0 on success
+     */
+    int whisper_set_mel(Pointer ctx, final float[] data, int n_len, int n_mel);
+    int whisper_set_mel_with_state(Pointer ctx, Pointer state, final float[] data, int n_len, int n_mel);
+
+    /**
+     * Run the Whisper encoder on the log mel spectrogram stored inside the default state in the provided whisper context.
+     * Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first.
+     * Offset can be used to specify the offset of the first frame in the spectrogram.
+     * @return 0 on success
+     */
+    int whisper_encode(Pointer ctx, int offset, int n_threads);
+
+    int whisper_encode_with_state(Pointer ctx, Pointer state, int offset, int n_threads);
+
+    /**
+     * Run the Whisper decoder to obtain the logits and probabilities for the next token.
+     * Make sure to call whisper_encode() first.
+     * tokens + n_tokens is the provided context for the decoder.
+     * n_past is the number of tokens to use from previous decoder calls.
+     * Returns 0 on success
+     * TODO: add support for multiple decoders
+     */
+    int whisper_decode(Pointer ctx, Pointer tokens, int n_tokens, int n_past, int n_threads);
+
+    /**
+     * @param ctx
+     * @param state
+     * @param tokens Pointer to int tokens
+     * @param n_tokens
+     * @param n_past
+     * @param n_threads
+     * @return
+     */
+    int whisper_decode_with_state(Pointer ctx, Pointer state, Pointer tokens, int n_tokens, int n_past, int n_threads);
+
+    /**
+     * Convert the provided text into tokens.
+     * The tokens pointer must be large enough to hold the resulting tokens.
+     * Returns the number of tokens on success, no more than n_max_tokens
+     * Returns -1 on failure
+     * TODO: not sure if correct
+     */
+    int whisper_tokenize(Pointer ctx, String text, Pointer tokens, int n_max_tokens);
+
+    /** Largest language id (i.e. number of available languages - 1) */
+    int whisper_lang_max_id();
+
+    /**
+     * @return the id of the specified language, returns -1 if not found.
+     * Examples:
+     *   "de" -> 2
+     *   "german" -> 2
+     */
+    int whisper_lang_id(String lang);
+
+    /** @return the short string of the specified language id (e.g. 2 -> "de"), returns nullptr if not found */
+    String whisper_lang_str(int id);
+
+    /**
+     * Use mel data at offset_ms to try and auto-detect the spoken language.
+     * Make sure to call whisper_pcm_to_mel() or whisper_set_mel() first
+     * Returns the top language id or negative on failure
+     * If not null, fills the lang_probs array with the probabilities of all languages
+     * The array must be whisper_lang_max_id() + 1 in size
+     *
+     * ref: https://github.com/openai/whisper/blob/main/whisper/decoding.py#L18-L69
+     */
+    int whisper_lang_auto_detect(Pointer ctx, int offset_ms, int n_threads, float[] lang_probs);
+
+    int whisper_lang_auto_detect_with_state(Pointer ctx, Pointer state, int offset_ms, int n_threads, float[] lang_probs);
+
+    int whisper_n_len           (Pointer ctx); // mel length
+    int whisper_n_len_from_state(Pointer state); // mel length
+    int whisper_n_vocab         (Pointer ctx);
+    int whisper_n_text_ctx      (Pointer ctx);
+    int whisper_n_audio_ctx     (Pointer ctx);
+    int whisper_is_multilingual (Pointer ctx);
+
+    int whisper_model_n_vocab      (Pointer ctx);
+    int whisper_model_n_audio_ctx  (Pointer ctx);
+    int whisper_model_n_audio_state(Pointer ctx);
+    int whisper_model_n_audio_head (Pointer ctx);
+    int whisper_model_n_audio_layer(Pointer ctx);
+    int whisper_model_n_text_ctx   (Pointer ctx);
+    int whisper_model_n_text_state (Pointer ctx);
+    int whisper_model_n_text_head  (Pointer ctx);
+    int whisper_model_n_text_layer (Pointer ctx);
+    int whisper_model_n_mels       (Pointer ctx);
+    int whisper_model_ftype        (Pointer ctx);
+    int whisper_model_type         (Pointer ctx);
+
+    /**
+     * Token logits obtained from the last call to whisper_decode().
+     * The logits for the last token are stored in the last row
+     * Rows: n_tokens
+     * Cols: n_vocab
+     */
+    float[] whisper_get_logits           (Pointer ctx);
+    float[] whisper_get_logits_from_state(Pointer state);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    String whisper_token_to_str(Pointer ctx, int token);
+    String whisper_model_type_readable(Pointer ctx);
+
+    // Special tokens
+    int whisper_token_eot (Pointer ctx);
+    int whisper_token_sot (Pointer ctx);
+    int whisper_token_prev(Pointer ctx);
+    int whisper_token_solm(Pointer ctx);
+    int whisper_token_not (Pointer ctx);
+    int whisper_token_beg (Pointer ctx);
+    int whisper_token_lang(Pointer ctx, int lang_id);
+
+    // Task tokens
+    int whisper_token_translate (Pointer ctx);
+    int whisper_token_transcribe(Pointer ctx);
+
+    // Performance information from the default state.
+    void whisper_print_timings(Pointer ctx);
+    void whisper_reset_timings(Pointer ctx);
+
+    // Note: Even if `whisper_full_params is stripped back to just 4 ints, JNA throws "Invalid memory access"
+    //       when `whisper_full_default_params()` tries to return a struct.
+    // WhisperFullParams whisper_full_default_params(int strategy);
+
+    /**
+     * Provides default params which can be used with `whisper_full()` etc.
+     * Because this function allocates memory for the params, the caller must call either:
+     * - call `whisper_free_params()`
+     * - `Native.free(Pointer.nativeValue(pointer));`
+     *
+     * @param strategy - WhisperSamplingStrategy.value
+     */
+    Pointer whisper_full_default_params_by_ref(int strategy);
+
+    void whisper_free_params(Pointer params);
+
+    /**
+     * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+     * Not thread safe for same context
+     * Uses the specified decoding strategy to obtain the text.
+     */
+    int whisper_full(Pointer ctx, WhisperFullParams params, final float[] samples, int n_samples);
+
+    int whisper_full_with_state(Pointer ctx, Pointer state, WhisperFullParams params, final float[] samples, int n_samples);
+
+    // Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+    // Result is stored in the default state of the context
+    // Not thread safe if executed in parallel on the same context.
+    // It seems this approach can offer some speedup in some cases.
+    // However, the transcription accuracy can be worse at the beginning and end of each chunk.
+    int whisper_full_parallel(Pointer ctx, WhisperFullParams params, final float[] samples, int n_samples, int n_processors);
+
+    /**
+     * Number of generated text segments.
+     * A segment can be a few words, a sentence, or even a paragraph.
+     * @param ctx Pointer to WhisperContext
+     */
+    int whisper_full_n_segments (Pointer ctx);
+
+    /**
+     * @param state Pointer to WhisperState
+     */
+    int whisper_full_n_segments_from_state(Pointer state);
+
+    /**
+     * Language id associated with the context's default state.
+     * @param ctx Pointer to WhisperContext
+     */
+    int whisper_full_lang_id(Pointer ctx);
+
+    /** Language id associated with the provided state */
+    int whisper_full_lang_id_from_state(Pointer state);
+
+    /**
+     * Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
+     * The resulting spectrogram is stored inside the default state of the provided whisper context.
+     * @return 0 on success
+     */
+    int whisper_pcm_to_mel_phase_vocoder(Pointer ctx, final float[] samples, int n_samples, int n_threads);
+
+    int whisper_pcm_to_mel_phase_vocoder_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);
+
+    /** Get the start time of the specified segment. */
+    long whisper_full_get_segment_t0(Pointer ctx, int i_segment);
+
+    /** Get the start time of the specified segment from the state. */
+    long whisper_full_get_segment_t0_from_state(Pointer state, int i_segment);
+
+    /** Get the end time of the specified segment. */
+    long whisper_full_get_segment_t1(Pointer ctx, int i_segment);
+
+    /** Get the end time of the specified segment from the state. */
+    long whisper_full_get_segment_t1_from_state(Pointer state, int i_segment);
+
+    /** Get the text of the specified segment. */
+    String whisper_full_get_segment_text(Pointer ctx, int i_segment);
+
+    /** Get the text of the specified segment from the state. */
+    String whisper_full_get_segment_text_from_state(Pointer state, int i_segment);
+
+    /** Get the number of tokens in the specified segment. */
+    int whisper_full_n_tokens(Pointer ctx, int i_segment);
+
+    /** Get the number of tokens in the specified segment from the state. */
+    int whisper_full_n_tokens_from_state(Pointer state, int i_segment);
+
+    /** Get the token text of the specified token in the specified segment. */
+    String whisper_full_get_token_text(Pointer ctx, int i_segment, int i_token);
+
+
+    /** Get the token text of the specified token in the specified segment from the state. */
+    String whisper_full_get_token_text_from_state(Pointer ctx, Pointer state, int i_segment, int i_token);
+
+    /** Get the token ID of the specified token in the specified segment. */
+    int whisper_full_get_token_id(Pointer ctx, int i_segment, int i_token);
+
+    /** Get the token ID of the specified token in the specified segment from the state. */
+    int whisper_full_get_token_id_from_state(Pointer state, int i_segment, int i_token);
+
+    /** Get token data for the specified token in the specified segment. */
+    WhisperTokenData whisper_full_get_token_data(Pointer ctx, int i_segment, int i_token);
+
+    /** Get token data for the specified token in the specified segment from the state. */
+    WhisperTokenData whisper_full_get_token_data_from_state(Pointer state, int i_segment, int i_token);
+
+    /** Get the probability of the specified token in the specified segment. */
+    float whisper_full_get_token_p(Pointer ctx, int i_segment, int i_token);
+
+    /** Get the probability of the specified token in the specified segment from the state. */
+    float whisper_full_get_token_p_from_state(Pointer state, int i_segment, int i_token);
+
+    /**
+     * Benchmark function for memcpy.
+     *
+     * @param nThreads Number of threads to use for the benchmark.
+     * @return The result of the benchmark.
+     */
+    int whisper_bench_memcpy(int nThreads);
+
+    /**
+     * Benchmark function for memcpy as a string.
+     *
+     * @param nThreads Number of threads to use for the benchmark.
+     * @return The result of the benchmark as a string.
+     */
+    String whisper_bench_memcpy_str(int nThreads);
+
+    /**
+     * Benchmark function for ggml_mul_mat.
+     *
+     * @param nThreads Number of threads to use for the benchmark.
+     * @return The result of the benchmark.
+     */
+    int whisper_bench_ggml_mul_mat(int nThreads);
+
+    /**
+     * Benchmark function for ggml_mul_mat as a string.
+     *
+     * @param nThreads Number of threads to use for the benchmark.
+     * @return The result of the benchmark as a string.
+     */
+    String whisper_bench_ggml_mul_mat_str(int nThreads);
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/bean/WhisperSegment.java

@@ -0,0 +1,47 @@
+package io.github.ggerganov.whispercpp.bean;
+
+/**
+ * Created by litonglinux@qq.com on 10/21/2023_7:48 AM
+ */
+public class WhisperSegment {
+  private long start, end;
+  private String sentence;
+
+  public WhisperSegment() {
+  }
+
+  public WhisperSegment(long start, long end, String sentence) {
+    this.start = start;
+    this.end = end;
+    this.sentence = sentence;
+  }
+
+  public long getStart() {
+    return start;
+  }
+
+  public long getEnd() {
+    return end;
+  }
+
+  public String getSentence() {
+    return sentence;
+  }
+
+  public void setStart(long start) {
+    this.start = start;
+  }
+
+  public void setEnd(long end) {
+    this.end = end;
+  }
+
+  public void setSentence(String sentence) {
+    this.sentence = sentence;
+  }
+
+  @Override
+  public String toString() {
+    return "[" + start + " --> " + end + "]:" + sentence;
+  }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/callbacks/WhisperEncoderBeginCallback.java

@@ -0,0 +1,24 @@
+package io.github.ggerganov.whispercpp.callbacks;
+
+import com.sun.jna.Callback;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.WhisperContext;
+import io.github.ggerganov.whispercpp.model.WhisperState;
+
+/**
+ * Callback before the encoder starts.
+ * If not null, called before the encoder starts.
+ * If it returns false, the computation is aborted.
+ */
+public interface WhisperEncoderBeginCallback extends Callback {
+
+    /**
+     * Callback method before the encoder starts.
+     *
+     * @param ctx        The whisper context.
+     * @param state      The whisper state.
+     * @param user_data  User data.
+     * @return True if the computation should proceed, false otherwise.
+     */
+    boolean callback(Pointer ctx, Pointer state, Pointer user_data);
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/callbacks/WhisperLogitsFilterCallback.java

@@ -0,0 +1,25 @@
+package io.github.ggerganov.whispercpp.callbacks;
+
+import com.sun.jna.Callback;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.model.WhisperTokenData;
+
+/**
+ * Callback to filter logits.
+ * Can be used to modify the logits before sampling.
+ * If not null, called after applying temperature to logits.
+ */
+public interface WhisperLogitsFilterCallback extends Callback {
+
+    /**
+     * Callback method to filter logits.
+     *
+     * @param ctx        The whisper context.
+     * @param state      The whisper state.
+     * @param tokens     The array of whisper_token_data.
+     * @param n_tokens   The number of tokens.
+     * @param logits     The array of logits.
+     * @param user_data  User data.
+     */
+    void callback(Pointer ctx, Pointer state, WhisperTokenData[] tokens, int n_tokens, float[] logits, Pointer user_data);
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/callbacks/WhisperNewSegmentCallback.java

@@ -0,0 +1,24 @@
+package io.github.ggerganov.whispercpp.callbacks;
+
+import com.sun.jna.Callback;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.WhisperContext;
+import io.github.ggerganov.whispercpp.model.WhisperState;
+
+/**
+ * Callback for the text segment.
+ * Called on every newly generated text segment.
+ * Use the whisper_full_...() functions to obtain the text segments.
+ */
+public interface WhisperNewSegmentCallback extends Callback {
+
+    /**
+     * Callback method for the text segment.
+     *
+     * @param ctx        The whisper context.
+     * @param state      The whisper state.
+     * @param n_new      The number of newly generated text segments.
+     * @param user_data  User data.
+     */
+    void callback(Pointer ctx, Pointer state, int n_new, Pointer user_data);
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/callbacks/WhisperProgressCallback.java

@@ -0,0 +1,22 @@
+package io.github.ggerganov.whispercpp.callbacks;
+
+import com.sun.jna.Callback;
+import com.sun.jna.Pointer;
+import io.github.ggerganov.whispercpp.WhisperContext;
+import io.github.ggerganov.whispercpp.model.WhisperState;
+
+/**
+ * Callback for progress updates.
+ */
+public interface WhisperProgressCallback extends Callback {
+
+    /**
+     * Callback method for progress updates.
+     *
+     * @param ctx        The whisper context.
+     * @param state      The whisper state.
+     * @param progress   The progress value.
+     * @param user_data  User data.
+     */
+    void callback(Pointer ctx, Pointer state, int progress, Pointer user_data);
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/ggml/GgmlTensor.java

@@ -0,0 +1,4 @@
+package io.github.ggerganov.whispercpp.ggml;
+
+public class GgmlTensor {
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/ggml/GgmlType.java

@@ -0,0 +1,18 @@
+package io.github.ggerganov.whispercpp.ggml;
+
+public enum GgmlType {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    REMOVED_GGML_TYPE_Q4_2,  // support has been removed
+    REMOVED_GGML_TYPE_Q4_3, // support has been removed
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q8_1,
+    GGML_TYPE_I8,
+    GGML_TYPE_I16,
+    GGML_TYPE_I32,
+    GGML_TYPE_COUNT,
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/model/EModel.java

@@ -0,0 +1,10 @@
+package io.github.ggerganov.whispercpp.model;
+
+public enum EModel {
+    MODEL_UNKNOWN,
+    MODEL_TINY,
+    MODEL_BASE,
+    MODEL_SMALL,
+    MODEL_MEDIUM,
+    MODEL_LARGE,
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/model/WhisperModel.java

@@ -0,0 +1,49 @@
+package io.github.ggerganov.whispercpp;
+
+import io.github.ggerganov.whispercpp.ggml.GgmlTensor;
+import io.github.ggerganov.whispercpp.model.EModel;
+
+public class WhisperModel {
+//    EModel type = EModel.MODEL_UNKNOWN;
+//
+//    WhisperHParams hparams;
+//    WhisperFilters filters;
+//
+//    // encoder.positional_embedding
+//    GgmlTensor e_pe;
+//
+//    // encoder.conv1
+//    GgmlTensor e_conv_1_w;
+//    GgmlTensor e_conv_1_b;
+//
+//    // encoder.conv2
+//    GgmlTensor e_conv_2_w;
+//    GgmlTensor e_conv_2_b;
+//
+//    // encoder.ln_post
+//    GgmlTensor e_ln_w;
+//    GgmlTensor e_ln_b;
+//
+//    // decoder.positional_embedding
+//    GgmlTensor d_pe;
+//
+//    // decoder.token_embedding
+//    GgmlTensor d_te;
+//
+//    // decoder.ln
+//    GgmlTensor d_ln_w;
+//    GgmlTensor d_ln_b;
+//
+//    std::vector<whisper_layer_encoder> layers_encoder;
+//    std::vector<whisper_layer_decoder> layers_decoder;
+//
+//    // context
+//    struct ggml_context * ctx;
+//
+//    // the model memory buffer is read-only and can be shared between processors
+//    std::vector<uint8_t> * buf;
+//
+//    // tensors
+//    int n_loaded;
+//    Map<String, GgmlTensor> tensors;
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/model/WhisperModelLoader.java

@@ -0,0 +1,62 @@
+package io.github.ggerganov.whispercpp.model;
+
+import com.sun.jna.Callback;
+import com.sun.jna.Pointer;
+import com.sun.jna.Structure;
+
+
+public class WhisperModelLoader extends Structure {
+    public Pointer context;
+    public ReadFunction read;
+    public EOFFunction eof;
+    public CloseFunction close;
+
+    public static class ReadFunction implements Callback {
+        public Pointer invoke(Pointer ctx, Pointer output, int readSize) {
+            // TODO
+            return ctx;
+        }
+    }
+
+    public static class EOFFunction implements Callback {
+        public boolean invoke(Pointer ctx) {
+            // TODO
+            return false;
+        }
+    }
+
+    public static class CloseFunction implements Callback {
+        public void invoke(Pointer ctx) {
+            // TODO
+        }
+    }
+
+//    public WhisperModelLoader(Pointer p) {
+//        super(p);
+//        read = new ReadFunction();
+//        eof = new EOFFunction();
+//        close = new CloseFunction();
+//        read.setCallback(this);
+//        eof.setCallback(this);
+//        close.setCallback(this);
+//        read.write();
+//        eof.write();
+//        close.write();
+//    }
+
+    public WhisperModelLoader() {
+        super();
+    }
+
+    public interface ReadCallback extends Callback {
+        Pointer invoke(Pointer ctx, Pointer output, int readSize);
+    }
+
+    public interface EOFCallback extends Callback {
+        boolean invoke(Pointer ctx);
+    }
+
+    public interface CloseCallback extends Callback {
+        void invoke(Pointer ctx);
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/model/WhisperState.java

@@ -0,0 +1,4 @@
+package io.github.ggerganov.whispercpp.model;
+
+public class WhisperState {
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/model/WhisperTokenData.java

@@ -0,0 +1,50 @@
+package io.github.ggerganov.whispercpp.model;
+
+import com.sun.jna.Structure;
+
+import java.util.Arrays;
+import java.util.List;
+
+/**
+ * Structure representing token data.
+ */
+public class WhisperTokenData extends Structure {
+
+    /** Token ID. */
+    public int id;
+
+    /** Forced timestamp token ID. */
+    public int tid;
+
+    /** Probability of the token. */
+    public float p;
+
+    /** Log probability of the token. */
+    public float plog;
+
+    /** Probability of the timestamp token. */
+    public float pt;
+
+    /** Sum of probabilities of all timestamp tokens. */
+    public float ptsum;
+
+    /**
+     * Start time of the token (token-level timestamp data).
+     * Do not use if you haven't computed token-level timestamps.
+     */
+    public long t0;
+
+    /**
+     * End time of the token (token-level timestamp data).
+     * Do not use if you haven't computed token-level timestamps.
+     */
+    public long t1;
+
+    /** Voice length of the token. */
+    public float vlen;
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("id", "tid", "p", "plog", "pt", "ptsum", "t0", "t1", "vlen");
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/BeamSearchParams.java

@@ -0,0 +1,19 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.Structure;
+
+import java.util.Arrays;
+import java.util.List;
+
+public class BeamSearchParams extends Structure {
+    /** ref: <a href="https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L265">...</a> */
+    public int beam_size;
+
+    /** ref: <a href="https://arxiv.org/pdf/2204.05424.pdf">...</a> */
+    public float patience;
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("beam_size", "patience");
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/CBool.java

@@ -0,0 +1,30 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.IntegerType;
+
+import java.util.function.BooleanSupplier;
+
+public class CBool extends IntegerType implements BooleanSupplier {
+    public static final int SIZE = 1;
+    public static final CBool FALSE = new CBool(0);
+    public static final CBool TRUE = new CBool(1);
+
+
+    public CBool() {
+        this(0);
+    }
+
+    public CBool(long value) {
+        super(SIZE, value, true);
+    }
+
+    @Override
+    public boolean getAsBoolean() {
+        return intValue() == 1;
+    }
+
+    @Override
+    public String toString() {
+        return intValue() == 1 ? "true" : "false";
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/GreedyParams.java

@@ -0,0 +1,16 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.Structure;
+
+import java.util.Collections;
+import java.util.List;
+
+public class GreedyParams extends Structure {
+    /** <a href="https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264">...</a> */
+    public int best_of;
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Collections.singletonList("best_of");
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperContextParams.java

@@ -0,0 +1,31 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.*;
+
+import java.util.Arrays;
+import java.util.List;
+
+/**
+ * Parameters for the whisper_init_from_file_with_params() function.
+ * If you change the order or add new parameters, make sure to update the default values in whisper.cpp:
+ * whisper_context_default_params()
+ */
+public class WhisperContextParams extends Structure {
+
+    public WhisperContextParams(Pointer p) {
+        super(p);
+    }
+
+    /** Use GPU for inference Number (default = true) */
+    public CBool use_gpu;
+
+    /** Use GPU for inference Number (default = true) */
+    public void useGpu(boolean enable) {
+        use_gpu = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("use_gpu");
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFilters.java

@@ -0,0 +1,10 @@
+package io.github.ggerganov.whispercpp.params;
+
+import java.util.List;
+
+public class WhisperFilters {
+    int n_mel;
+    int n_fft;
+
+    List<Float> data;
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java

@@ -0,0 +1,331 @@
+package io.github.ggerganov.whispercpp.params;
+
+import com.sun.jna.*;
+import io.github.ggerganov.whispercpp.callbacks.WhisperEncoderBeginCallback;
+import io.github.ggerganov.whispercpp.callbacks.WhisperLogitsFilterCallback;
+import io.github.ggerganov.whispercpp.callbacks.WhisperNewSegmentCallback;
+import io.github.ggerganov.whispercpp.callbacks.WhisperProgressCallback;
+
+import java.util.Arrays;
+import java.util.List;
+
+/**
+ * Parameters for the whisper_full() function.
+ * If you change the order or add new parameters, make sure to update the default values in whisper.cpp:
+ * whisper_full_default_params()
+ */
+public class WhisperFullParams extends Structure {
+
+    public WhisperFullParams(Pointer p) {
+        super(p);
+//        super(p, ALIGN_MSVC);
+//        super(p, ALIGN_GNUC);
+    }
+
+    /** Sampling strategy for whisper_full() function. */
+    public int strategy;
+
+    /** Number of threads. (default = 4) */
+    public int n_threads;
+
+    /** Maximum tokens to use from past text as a prompt for the decoder. (default = 16384) */
+    public int n_max_text_ctx;
+
+    /** Start offset in milliseconds. (default = 0) */
+    public int offset_ms;
+
+    /** Audio duration to process in milliseconds. (default = 0) */
+    public int duration_ms;
+
+    /** Translate flag. (default = false) */
+    public CBool translate;
+
+    /** The compliment of translateMode() */
+    public void transcribeMode() {
+        translate = CBool.FALSE;
+    }
+
+    /** The compliment of transcribeMode() */
+    public void translateMode() {
+        translate = CBool.TRUE;
+    }
+
+    /** Flag to indicate whether to use past transcription (if any) as an initial prompt for the decoder. (default = true) */
+    public CBool no_context;
+
+    /** Flag to indicate whether to use past transcription (if any) as an initial prompt for the decoder. (default = true) */
+    public void enableContext(boolean enable) {
+        no_context = enable ? CBool.FALSE : CBool.TRUE;
+    }
+
+    /** Generate timestamps or not? */
+    public CBool no_timestamps;
+
+    /** Flag to force single segment output (useful for streaming). (default = false) */
+    public CBool single_segment;
+
+    /** Flag to force single segment output (useful for streaming). (default = false) */
+    public void singleSegment(boolean single) {
+        single_segment = single ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
+    public CBool print_special;
+
+    /** Flag to print special tokens (e.g., <SOT>, <EOT>, <BEG>, etc.). (default = false) */
+    public void printSpecial(boolean enable) {
+        print_special = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Flag to print progress information. (default = true) */
+    public CBool print_progress;
+
+    /** Flag to print progress information. (default = true) */
+    public void printProgress(boolean enable) {
+        print_progress = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Flag to print results from within whisper.cpp (avoid it, use callback instead). (default = true) */
+    public CBool print_realtime;
+
+    /** Flag to print results from within whisper.cpp (avoid it, use callback instead). (default = true) */
+    public void printRealtime(boolean enable) {
+        print_realtime = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Flag to print timestamps for each text segment when printing realtime. (default = true) */
+    public CBool print_timestamps;
+
+    /** Flag to print timestamps for each text segment when printing realtime. (default = true) */
+    public void printTimestamps(boolean enable) {
+        print_timestamps = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** [EXPERIMENTAL] Flag to enable token-level timestamps. (default = false) */
+    public CBool token_timestamps;
+
+    /** [EXPERIMENTAL] Flag to enable token-level timestamps. (default = false) */
+    public void tokenTimestamps(boolean enable) {
+        token_timestamps = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** [EXPERIMENTAL] Timestamp token probability threshold (~0.01). (default = 0.01) */
+    public float thold_pt;
+
+    /** [EXPERIMENTAL] Timestamp token sum probability threshold (~0.01). */
+    public float thold_ptsum;
+
+    /** Maximum segment length in characters. (default = 0) */
+    public int max_len;
+
+    /** Flag to split on word rather than on token (when used with max_len). (default = false) */
+    public CBool split_on_word;
+
+    /** Flag to split on word rather than on token (when used with max_len). (default = false) */
+    public void splitOnWord(boolean enable) {
+        split_on_word = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Maximum tokens per segment (0, default = no limit) */
+    public int max_tokens;
+
+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public CBool speed_up;
+
+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public void speedUp(boolean enable) {
+        speed_up = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Overwrite the audio context size (0 = use default). */
+    public int audio_ctx;
+
+    /** Enable tinydiarize (default = false) */
+    public CBool tdrz_enable;
+
+    /** Enable tinydiarize (default = false) */
+    public void tdrzEnable(boolean enable) {
+        tdrz_enable = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Tokens to provide to the whisper decoder as an initial prompt.
+     * These are prepended to any existing text context from a previous call. */
+    public String initial_prompt;
+
+    /** Prompt tokens. (int*) */
+    public Pointer prompt_tokens;
+
+    public void setPromptTokens(int[] tokens) {
+        Memory mem = new Memory(tokens.length * 4L);
+        mem.write(0, tokens, 0, tokens.length);
+        prompt_tokens = mem;
+    }
+
+    /** Number of prompt tokens. */
+    public int prompt_n_tokens;
+
+    /** Language for auto-detection.
+     * For auto-detection, set to `null`, `""`, or "auto". */
+    public String language;
+
+    /** Flag to indicate whether to detect language automatically. */
+    public CBool detect_language;
+
+    /** Flag to indicate whether to detect language automatically. */
+    public void detectLanguage(boolean enable) {
+        detect_language = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    // Common decoding parameters.
+
+    /** Flag to suppress blank tokens. */
+    public CBool suppress_blank;
+
+    public void suppressBlanks(boolean enable) {
+        suppress_blank = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Flag to suppress non-speech tokens. */
+    public CBool suppress_non_speech_tokens;
+
+    /** Flag to suppress non-speech tokens. */
+    public void suppressNonSpeechTokens(boolean enable) {
+        suppress_non_speech_tokens = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
+    /** Initial decoding temperature. */
+    public float temperature;
+
+    /** Maximum initial timestamp. */
+    public float max_initial_ts;
+
+    /** Length penalty. */
+    public float length_penalty;
+
+    // Fallback parameters.
+
+    /** Temperature increment. */
+    public float temperature_inc;
+
+    /** Entropy threshold (similar to OpenAI's "compression_ratio_threshold"). */
+    public float entropy_thold;
+
+    /** Log probability threshold. */
+    public float logprob_thold;
+
+    /** No speech threshold. */
+    public float no_speech_thold;
+
+    /** Greedy decoding parameters. */
+    public GreedyParams greedy;
+
+    /**
+     * Beam search decoding parameters.
+     */
+    public BeamSearchParams beam_search;
+
+    public void setBestOf(int bestOf) {
+        if (greedy == null) {
+            greedy = new GreedyParams();
+        }
+        greedy.best_of = bestOf;
+    }
+
+    public void setBeamSize(int beamSize) {
+        if (beam_search == null) {
+            beam_search = new BeamSearchParams();
+        }
+        beam_search.beam_size = beamSize;
+    }
+
+    public void setBeamSizeAndPatience(int beamSize, float patience) {
+        if (beam_search == null) {
+            beam_search = new BeamSearchParams();
+        }
+        beam_search.beam_size = beamSize;
+        beam_search.patience = patience;
+    }
+
+    /**
+     * Callback for every newly generated text segment.
+     * WhisperNewSegmentCallback
+     */
+    public Pointer new_segment_callback;
+
+    /**
+     * User data for the new_segment_callback.
+     */
+    public Pointer new_segment_callback_user_data;
+
+    /**
+     * Callback on each progress update.
+     * WhisperProgressCallback
+     */
+    public Pointer progress_callback;
+
+    /**
+     * User data for the progress_callback.
+     */
+    public Pointer progress_callback_user_data;
+
+    /**
+     * Callback each time before the encoder starts.
+     * WhisperEncoderBeginCallback
+     */
+    public Pointer encoder_begin_callback;
+
+    /**
+     * User data for the encoder_begin_callback.
+     */
+    public Pointer encoder_begin_callback_user_data;
+
+    /**
+     * Callback by each decoder to filter obtained logits.
+     * WhisperLogitsFilterCallback
+     */
+    public Pointer logits_filter_callback;
+
+    /**
+     * User data for the logits_filter_callback.
+     */
+    public Pointer logits_filter_callback_user_data;
+
+
+    public void setNewSegmentCallback(WhisperNewSegmentCallback callback) {
+        new_segment_callback = CallbackReference.getFunctionPointer(callback);
+    }
+
+    public void setProgressCallback(WhisperProgressCallback callback) {
+        progress_callback = CallbackReference.getFunctionPointer(callback);
+    }
+
+    public void setEncoderBeginCallbackeginCallbackCallback(WhisperEncoderBeginCallback callback) {
+        encoder_begin_callback = CallbackReference.getFunctionPointer(callback);
+    }
+
+    public void setLogitsFilterCallback(WhisperLogitsFilterCallback callback) {
+        logits_filter_callback = CallbackReference.getFunctionPointer(callback);
+    }
+
+    /** Grammar stuff */
+    public Pointer grammar_rules;
+    public long n_grammar_rules;
+    public long i_start_rule;
+    public float grammar_penalty;
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
+                "no_context", "single_segment", "no_timestamps",
+                "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "tdrz_enable", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
+                "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
+                "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",
+                "new_segment_callback", "new_segment_callback_user_data",
+                "progress_callback", "progress_callback_user_data",
+                "encoder_begin_callback", "encoder_begin_callback_user_data",
+                "logits_filter_callback", "logits_filter_callback_user_data",
+                "grammar_rules", "n_grammar_rules", "i_start_rule", "grammar_penalty");
+    }
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperHParams.java

@@ -0,0 +1,15 @@
+package io.github.ggerganov.whispercpp.params;
+
+public class WhisperHParams {
+    int n_vocab       = 51864;
+    int n_audio_ctx   = 1500;
+    int n_audio_state = 384;
+    int n_audio_head  = 6;
+    int n_audio_layer = 4;
+    int n_text_ctx    = 448;
+    int n_text_state  = 384;
+    int n_text_head   = 6;
+    int n_text_layer  = 4;
+    int n_mels        = 80;
+    int ftype         = 1;
+}

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperSamplingStrategy.java

@@ -0,0 +1,10 @@
+package io.github.ggerganov.whispercpp.params;
+
+/** Available sampling strategies */
+public enum WhisperSamplingStrategy {
+    /** similar to OpenAI's GreedyDecoder */
+    WHISPER_SAMPLING_GREEDY,
+
+    /** similar to OpenAI's BeamSearchDecoder */
+    WHISPER_SAMPLING_BEAM_SEARCH
+}

bindings/java/src/test/java/io/github/ggerganov/whispercpp/WhisperCppTest.java

@@ -0,0 +1,144 @@
+package io.github.ggerganov.whispercpp;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+import io.github.ggerganov.whispercpp.bean.WhisperSegment;
+import io.github.ggerganov.whispercpp.params.CBool;
+import io.github.ggerganov.whispercpp.params.WhisperFullParams;
+import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
+import org.junit.jupiter.api.BeforeAll;
+import org.junit.jupiter.api.Test;
+import javax.sound.sampled.AudioInputStream;
+import javax.sound.sampled.AudioSystem;
+import java.io.File;
+import java.io.FileNotFoundException;
+import java.util.List;
+
+class WhisperCppTest {
+    private static WhisperCpp whisper = new WhisperCpp();
+    private static boolean modelInitialised = false;
+
+    @BeforeAll
+    static void init() throws FileNotFoundException {
+        // By default, models are loaded from ~/.cache/whisper/ and are usually named "ggml-${name}.bin"
+        // or you can provide the absolute path to the model file.
+        //String modelName = "../../models/ggml-tiny.bin";
+        String modelName = "../../models/ggml-tiny.en.bin";
+        try {
+            whisper.initContext(modelName);
+            //whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+            //whisper.getJavaDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+            modelInitialised = true;
+        } catch (FileNotFoundException ex) {
+            System.out.println("Model " + modelName + " not found");
+        }
+    }
+
+    @Test
+    void testGetDefaultFullParams_BeamSearch() {
+        // When
+        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+
+        // Then
+        assertEquals(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH.ordinal(), params.strategy);
+        assertNotEquals(0, params.n_threads);
+        assertEquals(16384, params.n_max_text_ctx);
+        assertFalse(params.translate);
+        assertEquals(0.01f, params.thold_pt);
+        assertEquals(5, params.beam_search.beam_size);
+        assertEquals(-1.0f, params.beam_search.patience);
+    }
+
+    @Test
+    void testGetDefaultFullParams_Greedy() {
+        // When
+        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+
+        // Then
+        assertEquals(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY.ordinal(), params.strategy);
+        assertNotEquals(0, params.n_threads);
+        assertEquals(16384, params.n_max_text_ctx);
+        assertEquals(5, params.greedy.best_of);
+    }
+
+    @Test
+    void testFullTranscribe() throws Exception {
+        if (!modelInitialised) {
+            System.out.println("Model not initialised, skipping test");
+            return;
+        }
+
+        // Given
+        File file = new File(System.getProperty("user.dir"), "../../samples/jfk.wav");
+        AudioInputStream audioInputStream = AudioSystem.getAudioInputStream(file);
+
+        byte[] b = new byte[audioInputStream.available()];
+        float[] floats = new float[b.length / 2];
+
+        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+        params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
+        params.print_progress = CBool.FALSE;
+        //params.initial_prompt = "and so my fellow Americans um, like";
+
+
+        try {
+            audioInputStream.read(b);
+
+            for (int i = 0, j = 0; i < b.length; i += 2, j++) {
+                int intSample = (int) (b[i + 1]) << 8 | (int) (b[i]) & 0xFF;
+                floats[j] = intSample / 32767.0f;
+            }
+
+            // When
+            String result = whisper.fullTranscribe(params, floats);
+
+            // Then
+            System.err.println(result);
+            assertEquals("And so my fellow Americans ask not what your country can do for you " +
+                    "ask what you can do for your country.",
+                    result.replace(",", ""));
+        } finally {
+            audioInputStream.close();
+        }
+    }
+
+    @Test
+    void testFullTranscribeWithTime() throws Exception {
+        if (!modelInitialised) {
+            System.out.println("Model not initialised, skipping test");
+            return;
+        }
+
+        // Given
+        File file = new File(System.getProperty("user.dir"), "../../samples/jfk.wav");
+        AudioInputStream audioInputStream = AudioSystem.getAudioInputStream(file);
+
+        byte[] b = new byte[audioInputStream.available()];
+        float[] floats = new float[b.length / 2];
+
+        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+        params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
+        params.print_progress = CBool.FALSE;
+        //params.initial_prompt = "and so my fellow Americans um, like";
+
+        try {
+            audioInputStream.read(b);
+
+            for (int i = 0, j = 0; i < b.length; i += 2, j++) {
+                int intSample = (int) (b[i + 1]) << 8 | (int) (b[i]) & 0xFF;
+                floats[j] = intSample / 32767.0f;
+            }
+
+            List<WhisperSegment> segments = whisper.fullTranscribeWithTime(params, floats);
+            assertTrue(segments.size() > 0, "The size of segments should be greater than 0");
+            for (WhisperSegment segment : segments) {
+                System.out.println(segment);
+            }
+        } finally {
+            audioInputStream.close();
+        }
+    }
+
+}

bindings/java/src/test/java/io/github/ggerganov/whispercpp/WhisperJnaLibraryTest.java

@@ -0,0 +1,17 @@
+package io.github.ggerganov.whispercpp;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+import org.junit.jupiter.api.Test;
+
+class WhisperJnaLibraryTest {
+
+    @Test
+    void testWhisperPrint_system_info() {
+        String systemInfo = WhisperCppJnaLibrary.instance.whisper_print_system_info();
+        // eg: "AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0
+        //    | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | COREML = 0 | "
+        System.out.println("System info: " + systemInfo);
+        assertTrue(systemInfo.length() > 10);
+    }
+}

bindings/javascript/emscripten.cpp

@@ -20,7 +20,7 @@ struct whisper_context * g_context;
 EMSCRIPTEN_BINDINGS(whisper) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         if (g_context == nullptr) {
-            g_context = whisper_init_from_file(path_model.c_str());
+            g_context = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
             if (g_context != nullptr) {
                 return true;
             } else {

bindings/javascript/libwhisper.worker.js

@@ -1 +1 @@
-"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:function(f){(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f)},postMessage:function(msg){parentPort.postMessage(msg)},performance:global.performance||{now:function(){return Date.now()}}})}var initializedJS=false;var pendingNotifiedProxyingQueues=[];function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module).then(function(instance){Module=instance})}else if(e.data.cmd==="run"){Module["__performance_now_clock_drift"]=performance.now()-e.data.time;Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};
+"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:f=>(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f),postMessage:msg=>parentPort.postMessage(msg),performance:global.performance||{now:Date.now}})}var initializedJS=false;function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"\n");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var module=Module["wasmModule"];Module["wasmModule"]=null;var instance=new WebAssembly.Instance(module,info);return receiveInstance(instance)};self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let messageQueue=[];self.onmessage=e=>messageQueue.push(e);self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"});for(let msg of messageQueue){handleMessage(msg)}self.onmessage=handleMessage};Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=(...args)=>{postMessage({cmd:"callHandler",handler:handler,args:args})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}whisper_factory(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["__emscripten_thread_mailbox_await"](e.data.pthread_ptr);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){Module["__embind_initialize_bindings"]();initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){throw ex}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="checkMailbox"){if(initializedJS){Module["checkMailbox"]()}}else if(e.data.cmd){err(`worker.js received unknown command ${e.data.cmd}`);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}}self.onmessage=handleMessage;

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.4.2",
+  "version": "1.5.0",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/.gitignore

@@ -1,6 +1,8 @@
 Makefile
 ggml.c
 ggml.h
+ggml-alloc.c
+ggml-alloc.h
 whisper.bundle
 whisper.cpp
 whisper.h

bindings/ruby/ext/extconf.rb

@@ -3,6 +3,14 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
+system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
 
 

bindings/ruby/ext/ggml-backend-impl.h

@@ -0,0 +1,87 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    typedef void * ggml_backend_buffer_context_t;
+
+    struct ggml_backend_buffer_i {
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
+        ggml_backend_buffer_context_t context;
+
+        size_t size;
+    };
+
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
+            struct ggml_backend_buffer_i           iface,
+                   ggml_backend_buffer_context_t   context,
+                   size_t                          size);
+
+    //
+    // Backend
+    //
+
+    typedef void * ggml_backend_context_t;
+
+    struct ggml_backend_i {
+        const char * (*get_name)(ggml_backend_t backend);
+
+        void (*free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
+
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);
+
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);
+
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // compute graph with a plan
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // check if the backend supports an operation
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+    };
+
+    struct ggml_backend {
+        struct ggml_backend_i iface;
+
+        ggml_backend_context_t context;
+    };
+
+#ifdef  __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-backend.c

@@ -0,0 +1,950 @@
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define UNUSED GGML_UNUSED
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+        struct ggml_backend                  * backend,
+        struct ggml_backend_buffer_i           iface,
+               ggml_backend_buffer_context_t   context,
+               size_t                          size) {
+    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
+
+    GGML_ASSERT(iface.get_base != NULL);
+
+    (*buffer) = (struct ggml_backend_buffer) {
+        /* .interface = */ iface,
+        /* .backend   = */ backend,
+        /* .context   = */ context,
+        /* .size      = */ size,
+    };
+
+    return buffer;
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    free(buffer);
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_get_alignment(buffer->backend);
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buffer->iface.get_alloc_size) {
+        return buffer->iface.get_alloc_size(buffer, tensor);
+    }
+    return ggml_nbytes(tensor);
+}
+
+void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // init_tensor is optional
+    if (buffer->iface.init_tensor) {
+        buffer->iface.init_tensor(buffer, tensor);
+    }
+}
+
+void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // free_tensor is optional
+    if (buffer->iface.free_tensor) {
+        buffer->iface.free_tensor(buffer, tensor);
+    }
+}
+
+// backend
+
+ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
+    return tensor->buffer ? tensor->buffer->backend : NULL;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return backend->iface.alloc_buffer(backend, size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return backend->iface.get_alignment(backend);
+}
+
+void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_t backend = ggml_get_backend(tensor);
+
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(backend != NULL && "tensor backend not set");
+
+    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    backend->iface.synchronize(backend);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_compute(backend, plan);
+}
+
+void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return backend->iface.supports_op(backend, op);
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
+    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
+
+    if (src == dst) {
+        return;
+    }
+
+    // TODO: allow backends to support copy to/from same backend
+
+    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
+        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
+    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
+        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
+    } else {
+        // shouldn't be hit when copying from/to CPU
+        #ifndef NDEBUG
+        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
+        #endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+// backend CPU
+
+struct ggml_backend_cpu_context {
+    int n_threads;
+    void * work_data;
+    size_t work_size;
+};
+
+static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
+    return "CPU";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    free(cpu_ctx->work_data);
+    free(cpu_ctx);
+    free(backend);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *)buffer->context;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+    UNUSED(buffer);
+}
+
+static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
+    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL, // no initialization required
+    /* .free_tensor    = */ NULL, // no cleanup required
+};
+
+// for buffers from ptr, free is not called
+static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
+    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
+    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
+    /* .init_tensor    = */ NULL,
+    /* .free_tensor    = */ NULL,
+};
+
+static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
+
+static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
+
+    GGML_ASSERT(data != NULL && "failed to allocate buffer");
+
+    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
+    return TENSOR_ALIGNMENT;
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy((char *)tensor->data + offset, data, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+
+    UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    cpu_plan->cgraph = *cgraph;
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
+    }
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    free(cpu_plan->cplan.work_data);
+    free(cpu_plan);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        // TODO: may be faster to free and use malloc to avoid the copy
+        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        cpu_ctx->work_size = cplan.work_size;
+    }
+
+    cplan.work_data = cpu_ctx->work_data;
+
+    ggml_graph_compute(cgraph, &cplan);
+}
+
+static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    return true;
+    UNUSED(backend);
+    UNUSED(op);
+}
+
+static struct ggml_backend_i cpu_backend_i = {
+    /* .get_name            = */ ggml_backend_cpu_name,
+    /* .free                = */ ggml_backend_cpu_free,
+    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
+    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
+    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
+    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
+    /* .synchronize         = */ ggml_backend_cpu_synchronize,
+    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
+    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
+    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op         = */ ggml_backend_cpu_supports_op,
+};
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
+
+    ctx->n_threads = GGML_DEFAULT_N_THREADS;
+    ctx->work_data = NULL;
+    ctx->work_size = 0;
+
+    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
+
+    *cpu_backend = (struct ggml_backend) {
+        /* .interface = */ cpu_backend_i,
+        /* .context   = */ ctx
+    };
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend->iface.get_name == ggml_backend_cpu_name;
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
+    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
+}
+
+// scheduler
+
+#define GGML_MAX_BACKENDS 4
+#define GGML_MAX_SPLITS 256
+#define GGML_MAX_SPLIT_INPUTS 16
+
+struct ggml_backend_sched_split {
+    ggml_tallocr_t tallocr;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    struct ggml_cgraph * graph;
+};
+
+struct ggml_backend_sched {
+    int n_backends;
+    ggml_backend_t backends[GGML_MAX_BACKENDS];
+    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
+
+    ggml_gallocr_t galloc;
+
+    struct ggml_hash_set    hash_set;
+    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
+    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
+
+    struct ggml_cgraph * graph;
+    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
+    int n_splits;
+
+    struct ggml_context * ctx;
+
+    // align context_buffer to GGML_MEM_ALIGN
+    #ifdef _MSC_VER
+    __declspec(align(GGML_MEM_ALIGN))
+    #else
+    __attribute__((aligned(GGML_MEM_ALIGN)))
+    #endif
+    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
+};
+
+#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
+#define node_allocr(node) sched->node_talloc[hash_id(node)]
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// returns the priority of the backend, lower is better
+static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->tallocs[i] == allocr) {
+            return i;
+        }
+    }
+    return INT_MAX;
+}
+
+// returns the backend that should be used for the node based on the current locations
+char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
+static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
+    // ie. kv cache updates
+    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
+    // dst
+    ggml_backend_t cur_backend = ggml_get_backend(node);
+    if (cur_backend != NULL) {
+        sprintf(causes[hash_id(node)], "1.dst");
+        return cur_backend;
+    }
+
+    // view_src
+    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
+        sprintf(causes[hash_id(node)], "1.vsrc");
+        return ggml_get_backend(node->view_src);
+    }
+
+    // src
+    int cur_prio = INT_MAX;
+    size_t cur_size = 0;
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = node->src[i];
+        if (src == NULL) {
+            break;
+        }
+        ggml_backend_t src_backend = ggml_get_backend(src);
+        if (src_backend != NULL) {
+            int src_prio = sched_backend_prio(sched, src_backend);
+            size_t src_size = ggml_nbytes(src);
+            if (src_prio < cur_prio && src_size >= cur_size) {
+                cur_prio = src_prio;
+                cur_size = src_size;
+                cur_backend = src_backend;
+                sprintf(causes[hash_id(node)], "1.src%d", i);
+            }
+        }
+    }
+    return cur_backend;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        sprintf(buffer, "%zuM", size/1024/1024);
+    } else {
+        sprintf(buffer, "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
+            ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
+            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs);
+            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
+            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+// TODO: merge passes
+static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset state
+    size_t hash_size = sched->hash_set.size;
+    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
+    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
+    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
+    sched->n_splits = 0;
+
+    struct ggml_init_params params = {
+        /*.mem_size =   */ sizeof(sched->context_buffer),
+        /*.mem_buffer = */ sched->context_buffer,
+        /*.no_alloc =   */ true
+    };
+
+    if (sched->ctx != NULL) {
+        ggml_free(sched->ctx);
+    }
+
+    sched->ctx = ggml_init(params);
+
+    // pass 1: assign backends to ops with allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        if (node_allocr(leaf) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
+        if (leaf_backend == NULL && leaf->view_src != NULL) {
+            leaf_backend = ggml_get_backend(leaf->view_src);
+        }
+        if (leaf_backend != NULL) {
+            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node_allocr(node) != NULL) {
+            // do not overwrite user assignments
+            continue;
+        }
+        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
+        if (node_backend != NULL) {
+            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
+        }
+    }
+    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 2: assign backends to ops from current assignments
+    // TODO:
+    //  - reuse sched_backend_from_cur
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            int    cur_prio = INT_MAX;
+            size_t cur_size = 0;
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    break;
+                }
+                ggml_tallocr_t src_allocr = node_allocr(src);
+                if (src_allocr != NULL) {
+                    int    src_prio = sched_allocr_prio(sched, src_allocr);
+                    size_t src_size = ggml_nbytes(src);
+                    if (src_prio < cur_prio && src_size >= cur_size) {
+                        cur_prio = src_prio;
+                        cur_size = src_size;
+                        node_allocr = src_allocr;
+                        sprintf(causes[hash_id(node)], "2.src%d", j);
+                    }
+                }
+            }
+            if (node_allocr != NULL) {
+                node_allocr(node) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 3: assign backends to remaining src from dst (should only be leafs)
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr == NULL) {
+                node_allocr(src) = node_allocr;
+            }
+        }
+    }
+    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+
+    // pass 4: split graph, find tensors that need to be copied
+    // TODO:
+    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
+    // find first backend
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (node->view_src == NULL) {
+            sched->splits[0].tallocr = node_allocr(node);
+            break;
+        }
+    }
+    sched->splits[0].i_start = 0;
+    sched->splits[0].n_inputs = 0;
+    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
+    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
+    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+
+        ggml_tallocr_t node_allocr = node_allocr(node);
+
+        if (node_allocr != cur_allocr) {
+            sched->splits[cur_split].i_end = i;
+            cur_split++;
+            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
+            sched->splits[cur_split].tallocr = node_allocr;
+            sched->splits[cur_split].i_start = i;
+            sched->splits[cur_split].n_inputs = 0;
+            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
+            cur_allocr = node_allocr;
+            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
+        }
+
+        // find inputs that are not on the same backend
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr) {
+                int n_inputs = sched->splits[cur_split].n_inputs++;
+                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
+
+                // create copies
+                size_t id = hash_id(src);
+                if (sched->node_copies[id][cur_backend_id] == NULL) {
+                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                    sched->node_copies[id][cur_backend_id] = tensor_copy;
+                    node_allocr(tensor_copy) = cur_allocr;
+                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
+                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
+                }
+                node->src[j] = sched->node_copies[id][cur_backend_id];
+            }
+        }
+    }
+    sched->splits[cur_split].i_end = graph->n_nodes;
+    sched->n_splits = cur_split + 1;
+
+    //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout);
+
+#if 1
+    // sanity check: all sources should have the same backend as the node
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        ggml_tallocr_t node_allocr = node_allocr(node);
+        if (node_allocr == NULL) {
+            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                break;
+            }
+            ggml_tallocr_t src_allocr = node_allocr(src);
+            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
+                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
+                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
+                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
+            }
+        }
+    }
+#endif
+
+    // create copies of the graph for each split
+    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
+            input_cpy->src[0] = input;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+    sched->graph = graph_copy;
+}
+
+static void sched_alloc_splits(ggml_backend_sched_t sched) {
+    ggml_gallocr_alloc_graph_n(
+        sched->galloc,
+        sched->graph,
+        sched->hash_set,
+        sched->node_talloc);
+}
+
+static void sched_compute_splits(ggml_backend_sched_t sched) {
+    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
+    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
+
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
+        int split_backend_id = sched_backend_prio(sched, split_backend);
+
+        // copy the input tensors to the split backend
+        uint64_t copy_start_us = ggml_time_us();
+        for (int j = 0; j < split->n_inputs; j++) {
+            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
+            if (split->inputs[j]->buffer == NULL) {
+                if (split->inputs[j]->view_src == NULL) {
+                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
+                    exit(1);
+                }
+                struct ggml_tensor * view = split->inputs[j];
+                view->backend = view->view_src->backend;
+                view->buffer  = view->view_src->buffer;
+                view->data    = (char *)view->view_src->data + view->view_offs;
+                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
+            }
+            if (input_cpy->buffer == NULL) {
+                fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
+                exit(1);
+            }
+            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
+            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
+            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
+        }
+        // ggml_backend_synchronize(split_backend);
+        int64_t copy_end_us = ggml_time_us();
+        copy_us[split_backend_id] += copy_end_us - copy_start_us;
+
+#if 0
+        char split_filename[GGML_MAX_NAME];
+        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
+        ggml_graph_dump_dot(split->graph, NULL, split_filename);
+#endif
+
+        uint64_t compute_start_us = ggml_time_us();
+        ggml_backend_graph_compute(split_backend, split->graph);
+        // ggml_backend_synchronize(split_backend);
+        uint64_t compute_end_us = ggml_time_us();
+        compute_us[split_backend_id] += compute_end_us - compute_start_us;
+    }
+
+#if 0
+    // per-backend timings
+    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (copy_us[i] > 0 || compute_us[i] > 0) {
+            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
+        }
+    }
+#endif
+}
+
+static void sched_reset(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_reset(sched->tallocs[i]);
+    }
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
+    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
+
+    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
+    memset(sched, 0, sizeof(struct ggml_backend_sched));
+
+    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
+
+    sched->n_backends = n_backends;
+    for (int i = 0; i < n_backends; i++) {
+        sched->backends[i] = backends[i];
+    }
+
+    sched->galloc = ggml_gallocr_new();
+
+    // init measure allocs for each backend
+    for (int i = 0; i < n_backends; i++) {
+        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
+    }
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_tallocr_free(sched->tallocs[i]);
+    }
+    ggml_gallocr_free(sched->galloc);
+    free(sched->hash_set.keys);
+    free(sched->node_talloc);
+    free(sched->node_copies);
+    free(sched);
+}
+
+void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    // initialize hash tables
+    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
+    sched->hash_set.size = hash_size;
+    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
+    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
+    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
+
+    sched_split_graph(sched, measure_graph);
+    sched_alloc_splits(sched);
+
+    // allocate buffers and reset allocators
+    for (int i = 0; i < sched->n_backends; i++) {
+        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
+        ggml_tallocr_free(sched->tallocs[i]);
+        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
+    }
+
+    sched_reset(sched);
+}
+
+void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+
+    sched_split_graph(sched, graph);
+    sched_alloc_splits(sched);
+    sched_compute_splits(sched);
+    sched_reset(sched);
+}
+
+ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return sched->tallocs[backend_index];
+}
+
+ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
+}
+
+void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = sched_backend_prio(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    node_allocr(node) = sched->tallocs[backend_index];
+}

bindings/ruby/ext/ggml-backend.h

@@ -0,0 +1,136 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer
+    //
+
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+
+    // backend buffer functions
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+
+    //
+    // Backend
+    //
+
+    struct ggml_backend;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
+
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backends to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
+        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
+
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
+
+        // in build_graph:
+        build_graph(...) {
+            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
+            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
+            ggml_allocr_alloc(alloc_cpu, tensor);
+
+            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
+        }
+
+        // allocate backend buffers from measure graph
+        ggml_backend_sched_init_measure(sched, measure_graph);
+
+        // the scheduler is now ready to compute graphs
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+    */
+
+    struct ggml_backend_sched;
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+
+    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+
+    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+
+    // Allocate a graph on the backend scheduler
+    GGML_API void ggml_backend_sched_graph_compute(
+            ggml_backend_sched_t sched,
+            struct ggml_cgraph * graph);
+
+#ifdef  __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-impl.h

@@ -0,0 +1,249 @@
+#pragma once
+
+#include "ggml.h"
+
+// GGML internal header
+
+#include <assert.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#endif
+
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
+
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // __ARM_NEON
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+#endif
+
+#define GGML_HASHTABLE_FULL ((size_t)-1)
+#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
+
+bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
+size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+// return index, asserts if table is full
+size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
+
+#ifdef __cplusplus
+}
+#endif

bindings/ruby/ext/ggml-quants.h

@@ -0,0 +1,224 @@
+#pragma once
+
+#include "ggml-impl.h"
+
+// GGML internal header
+
+#include <stdint.h>
+#include <stddef.h>
+
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
+#endif
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
+// Quantization
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
+
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
+
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
+
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
+
+// Dequantization
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
+
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

bindings/ruby/ext/ruby_whisper.cpp

@@ -87,7 +87,7 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   if (!rb_respond_to(whisper_model_file_path, rb_intern("to_s"))) {
     rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
   }
-  rw->context = whisper_init_from_file(StringValueCStr(whisper_model_file_path));
+  rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), whisper_context_default_params());
   if (rw->context == nullptr) {
     rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
   }

coreml/whisper-encoder-impl.h

@@ -123,7 +123,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v
 
 /**
     Make a prediction using the convenience interface
-    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
+    @param logmel_data as 1 × n_mel × 3000 3-dimensional array of floats:
     @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
     @return the prediction as whisper_encoder_implOutput
 */

coreml/whisper-encoder.h

@@ -3,6 +3,8 @@
 // Code is derived from the work of Github user @wangchou
 // ref: https://github.com/wangchou/callCoreMLFromCpp
 
+#include <stdint.h>
+
 #if __cplusplus
 extern "C" {
 #endif
@@ -14,6 +16,8 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx);
 
 void whisper_coreml_encode(
         const whisper_coreml_context * ctx,
+                             int64_t   n_ctx,
+                             int64_t   n_mel,
                                float * mel,
                                float * out);
 

coreml/whisper-encoder.mm

@@ -22,7 +22,13 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
 
     NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
 
-    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+    // select which device to run the Core ML model on
+    MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
+    config.computeUnits = MLComputeUnitsCPUAndGPU;
+    //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
+    //config.computeUnits = MLComputeUnitsAll;
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 
     if (data == NULL) {
         return NULL;
@@ -42,20 +48,24 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx) {
 
 void whisper_coreml_encode(
         const whisper_coreml_context * ctx,
+                             int64_t   n_ctx,
+                             int64_t   n_mel,
                                float * mel,
                                float * out) {
     MLMultiArray * inMultiArray = [
         [MLMultiArray alloc] initWithDataPointer: mel
-                                           shape: @[@1, @80, @3000]
+                                           shape: @[@1, @(n_mel), @(n_ctx)]
                                         dataType: MLMultiArrayDataTypeFloat32
-                                         strides: @[@(240000), @(3000), @1]
+                                         strides: @[@(n_ctx*n_mel), @(n_ctx), @1]
                                      deallocator: nil
                                            error: nil
     ];
 
-    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+    @autoreleasepool {
+        whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
 
-    memcpy(out, outCoreML.output.dataPointer, outCoreML.output.count * sizeof(float));
+        memcpy(out, outCoreML.output.dataPointer, outCoreML.output.count * sizeof(float));
+    }
 }
 
 #if __cplusplus

examples/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(${TARGET} STATIC
     common.cpp
     common-ggml.h
     common-ggml.cpp
+    grammar-parser.cpp
     )
 
 include(DefaultTargetOptions)
@@ -64,9 +65,13 @@ elseif(CMAKE_JS_VERSION)
 else()
     add_subdirectory(main)
     add_subdirectory(stream)
+    add_subdirectory(server)
     add_subdirectory(command)
     add_subdirectory(bench)
     add_subdirectory(quantize)
     add_subdirectory(talk)
     add_subdirectory(talk-llama)
+    add_subdirectory(lsp)
 endif()
+
+add_subdirectory(wchess)

examples/addon.node/__test__/whisper.spec.js

@@ -11,6 +11,7 @@ const whisperParamsMock = {
   language: "en",
   model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
   fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+  use_gpu: true,
 };
 
 describe("Run whisper.node", () => {

examples/addon.node/addon.cpp

@@ -36,6 +36,7 @@ struct whisper_params {
     bool print_colors   = false;
     bool print_progress = false;
     bool no_timestamps  = false;
+    bool use_gpu        = true;
 
     std::string language = "en";
     std::string prompt;
@@ -153,7 +154,9 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
 
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     if (ctx == nullptr) {
         fprintf(stderr, "error: failed to initialize whisper context\n");
@@ -315,10 +318,12 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
   std::string language = whisper_params.Get("language").As<Napi::String>();
   std::string model = whisper_params.Get("model").As<Napi::String>();
   std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
 
   params.language = language;
   params.model = model;
   params.fname_inp.emplace_back(input);
+  params.use_gpu = use_gpu;
 
   Napi::Function callback = info[1].As<Napi::Function>();
   Worker* worker = new Worker(callback, params);

examples/addon.node/index.js

@@ -11,6 +11,7 @@ const whisperParams = {
   language: "en",
   model: path.join(__dirname, "../../models/ggml-base.en.bin"),
   fname_inp: "../../samples/jfk.wav",
+  use_gpu: true,
 };
 
 const arguments = process.argv.slice(2);

examples/bench.wasm/emscripten.cpp

@@ -23,7 +23,9 @@ void bench_main(size_t index) {
 
     fprintf(stderr, "%s: running benchmark with %d threads - please wait...\n", __func__, n_threads);
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return;
     }
@@ -57,7 +59,7 @@ EMSCRIPTEN_BINDINGS(bench) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     if (g_worker.joinable()) {
                         g_worker.join();

examples/bench/bench.cpp

@@ -1,6 +1,7 @@
 #include "whisper.h"
 
 #include <cstdio>
+#include <cstring>
 #include <string>
 #include <thread>
 
@@ -10,6 +11,8 @@ struct whisper_params {
     int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
 
     std::string model = "models/ggml-base.en.bin";
+
+    bool use_gpu = true;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -22,9 +25,10 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
-        else if (arg == "-m" || arg == "--model")   { params.model     = argv[++i]; }
-        else if (arg == "-w" || arg == "--what")    { params.what     = atoi(argv[++i]); }
+        else if (arg == "-t"  || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")    { params.what      = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")  { params.use_gpu   = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -44,16 +48,20 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
     fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
-    fprintf(stderr, "                           %-7s  0 - whisper encoder\n",                         "");
+    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
     fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
     fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int whisper_bench_encoder(const whisper_params & params) {
+int whisper_bench_full(const whisper_params & params) {
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     {
         fprintf(stderr, "\n");
@@ -65,16 +73,65 @@ int whisper_bench_encoder(const whisper_params & params) {
         return 2;
     }
 
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
         fprintf(stderr, "error: failed to set mel: %d\n", ret);
         return 3;
     }
-
+    // heat encoder
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
         return 4;
     }
 
+    whisper_token tokens[512];
+    memset(tokens, 0, sizeof(tokens));
+
+    // prompt heat
+    if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation heat
+    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    whisper_reset_timings(ctx);
+
+    // actual run
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // batched decoding
+    for (int i = 0; i < 64; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // prompt processing
+    for (int i = 0; i < 16; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
     whisper_print_timings(ctx);
     whisper_free(ctx);
 
@@ -103,7 +160,7 @@ int main(int argc, char ** argv) {
     int ret = -1;
 
     switch (params.what) {
-        case 0: ret = whisper_bench_encoder(params);                break;
+        case 0: ret = whisper_bench_full(params);                break;
         case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
         case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
         default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;

examples/command.wasm/emscripten.cpp

@@ -243,7 +243,7 @@ EMSCRIPTEN_BINDINGS(command) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/command/command.cpp

@@ -6,9 +6,10 @@
 // ref: https://github.com/ggerganov/whisper.cpp/issues/171
 //
 
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
+#include "grammar-parser.h"
 
 #include <sstream>
 #include <cassert>
@@ -21,6 +22,11 @@
 #include <vector>
 #include <map>
 
+bool file_exists(const std::string & fname) {
+    std::ifstream f(fname.c_str());
+    return f.good();
+}
+
 // command-line parameters
 struct whisper_params {
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -30,20 +36,27 @@ struct whisper_params {
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
 
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
+    float vad_thold  = 0.6f;
+    float freq_thold = 100.0f;
+
+    float grammar_penalty = 100.0f;
+
+    grammar_parser::parse_state grammar_parsed;
 
     bool speed_up      = false;
     bool translate     = false;
     bool print_special = false;
     bool print_energy  = false;
     bool no_timestamps = true;
+    bool use_gpu       = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
     std::string fname_out;
     std::string commands;
     std::string prompt;
+    std::string context;
+    std::string grammar;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -68,11 +81,15 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
         else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
         else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
         else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
         else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
+        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -101,21 +118,36 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
     fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
     fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
     fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
     fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
     fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
     fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
+    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
     fprintf(stderr, "\n");
 }
 
-std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
+std::string transcribe(
+                 whisper_context * ctx,
+            const whisper_params & params,
+        const std::vector<float> & pcmf32,
+               const std::string & grammar_rule,
+                           float & logprob_min,
+                           float & logprob_sum,
+                             int & n_tokens,
+                         int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
-    prob = 0.0f;
+    logprob_min = 0.0f;
+    logprob_sum = 0.0f;
+    n_tokens    = 0;
     t_ms = 0;
 
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    //whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH);
 
     wparams.print_progress   = false;
     wparams.print_special    = params.print_special;
@@ -123,19 +155,41 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     wparams.print_timestamps = !params.no_timestamps;
     wparams.translate        = params.translate;
     wparams.no_context       = true;
+    wparams.no_timestamps    = params.no_timestamps;
     wparams.single_segment   = true;
     wparams.max_tokens       = params.max_tokens;
     wparams.language         = params.language.c_str();
     wparams.n_threads        = params.n_threads;
 
-    wparams.audio_ctx        = params.audio_ctx;
-    wparams.speed_up         = params.speed_up;
+    wparams.audio_ctx = params.audio_ctx;
+    wparams.speed_up  = params.speed_up;
+
+    wparams.temperature     = 0.4f;
+    wparams.temperature_inc = 1.0f;
+    wparams.greedy.best_of  = 5;
+
+    wparams.beam_search.beam_size = 5;
+
+    wparams.initial_prompt = params.context.data();
+
+    const auto & grammar_parsed = params.grammar_parsed;
+    auto grammar_rules = grammar_parsed.c_rules();
+
+    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
+        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
+            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, grammar_rule.c_str());
+        } else {
+            wparams.grammar_rules   = grammar_rules.data();
+            wparams.n_grammar_rules = grammar_rules.size();
+            wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
+            wparams.grammar_penalty = params.grammar_penalty;
+        }
+    }
 
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         return "";
     }
 
-    int prob_n = 0;
     std::string result;
 
     const int n_segments = whisper_full_n_segments(ctx);
@@ -144,19 +198,17 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
 
         result += text;
 
-        const int n_tokens = whisper_full_n_tokens(ctx, i);
-        for (int j = 0; j < n_tokens; ++j) {
+        const int n = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n; ++j) {
             const auto token = whisper_full_get_token_data(ctx, i, j);
 
-            prob += token.p;
-            ++prob_n;
+            if(token.plog > 0.0f) exit(0);
+            logprob_min = std::min(logprob_min, token.plog);
+            logprob_sum += token.plog;
+            ++n_tokens;
         }
     }
 
-    if (prob_n > 0) {
-        prob /= prob_n;
-    }
-
     const auto t_end = std::chrono::high_resolution_clock::now();
     t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
 
@@ -247,7 +299,7 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
         fprintf(stderr, " ]\n");
     }
 
-    std::string  k_prompt = "select one from the available words: ";
+    std::string k_prompt = "select one from the available words: ";
     for (int i = 0; i < (int) allowed_commands.size(); ++i) {
         if (i > 0) {
             k_prompt += ", ";
@@ -415,7 +467,9 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
     bool is_running = true;
     bool ask_prompt = true;
 
-    float prob = 0.0f;
+    float logprob_min = 0.0f;
+    float logprob_sum = 0.0f;
+    int   n_tokens    = 0;
 
     std::vector<float> pcmf32_cur;
 
@@ -453,7 +507,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
                 // detect the commands
                 audio.get(params.command_ms, pcmf32_cur);
 
-                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "", logprob_min, logprob_sum, n_tokens, t_ms));
 
                 const auto words = get_words(txt);
 
@@ -489,18 +543,27 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
 
 // general-purpose mode
 // freely transcribe the voice into text
-int process_general_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
     bool is_running  = true;
     bool have_prompt = false;
     bool ask_prompt  = true;
 
-    float prob0 = 0.0f;
-    float prob  = 0.0f;
+    float logprob_min0 = 0.0f;
+    float logprob_min  = 0.0f;
+
+    float logprob_sum0 = 0.0f;
+    float logprob_sum  = 0.0f;
+
+    int n_tokens0 = 0;
+    int n_tokens  = 0;
 
     std::vector<float> pcmf32_cur;
     std::vector<float> pcmf32_prompt;
 
-    const std::string k_prompt = "Ok Whisper, start listening for commands.";
+    std::string k_prompt = "Ok Whisper, start listening for commands.";
+    if (!params.prompt.empty()) {
+        k_prompt = params.prompt;
+    }
 
     fprintf(stderr, "\n");
     fprintf(stderr, "%s: general-purpose mode\n", __func__);
@@ -533,9 +596,11 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                     // wait for activation phrase
                     audio.get(params.prompt_ms, pcmf32_cur);
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "prompt", logprob_min0, logprob_sum0, n_tokens0, t_ms));
 
-                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+                    const float p = 100.0f * std::exp(logprob_min0);
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms, p = %.2f%%)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms, p);
 
                     const float sim = similarity(txt, k_prompt);
 
@@ -556,19 +621,30 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                     // we have heard the activation phrase, now detect the commands
                     audio.get(params.command_ms, pcmf32_cur);
 
+                    //printf("len prompt:  %.4f\n", pcmf32_prompt.size() / (float) WHISPER_SAMPLE_RATE);
+                    //printf("len command: %.4f\n", pcmf32_cur.size() / (float) WHISPER_SAMPLE_RATE);
+
+                    // prepend 3 second of silence
+                    pcmf32_cur.insert(pcmf32_cur.begin(), 3.0f*WHISPER_SAMPLE_RATE, 0.0f);
+
                     // prepend the prompt audio
                     pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
 
-                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "root", logprob_min, logprob_sum, n_tokens, t_ms));
 
-                    prob = 100.0f*(prob - prob0);
+                    //const float p = 100.0f * std::exp((logprob - logprob0) / (n_tokens - n_tokens0));
+                    const float p = 100.0f * std::exp(logprob_min);
 
                     //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
 
                     // find the prompt in the text
                     float best_sim = 0.0f;
                     size_t best_len = 0;
-                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                    for (size_t n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        if (n >= txt.size()) {
+                            break;
+                        }
+
                         const auto prompt = txt.substr(0, n);
 
                         const float sim = similarity(prompt, k_prompt);
@@ -581,9 +657,16 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
                         }
                     }
 
-                    const std::string command = ::trim(txt.substr(best_len));
+                    fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
+                    if (best_len == 0) {
+                        fprintf(stdout, "%s: WARNING: command not recognized, try again\n", __func__);
+                    } else {
+                        // cut the prompt from the decoded text
+                        const std::string command = ::trim(txt.substr(best_len));
+
+                        fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    }
 
-                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
                     fprintf(stdout, "\n");
                 }
 
@@ -610,7 +693,10 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     // print some info about the processing
     {
@@ -648,12 +734,36 @@ int main(int argc, char ** argv) {
 
     int  ret_val = 0;
 
-    if (!params.commands.empty()) {
-        ret_val = process_command_list(ctx, audio, params);
-    } else if (!params.prompt.empty()) {
-        ret_val = always_prompt_transcription(ctx, audio, params);
-    } else {
-        ret_val = process_general_transcription(ctx, audio, params);
+    if (!params.grammar.empty()) {
+        auto & grammar = params.grammar_parsed;
+        if (file_exists(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            ret_val = 1;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (ret_val == 0) {
+        if (!params.commands.empty()) {
+            ret_val = process_command_list(ctx, audio, params);
+        } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
+            ret_val = always_prompt_transcription(ctx, audio, params);
+        } else {
+            ret_val = process_general_transcription(ctx, audio, params);
+        }
     }
 
     audio.pause();

examples/common-ggml.cpp

@@ -9,6 +9,11 @@ static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
     {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
     {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
     {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+    {"q2_k", GGML_FTYPE_MOSTLY_Q2_K},
+    {"q3_k", GGML_FTYPE_MOSTLY_Q3_K},
+    {"q4_k", GGML_FTYPE_MOSTLY_Q4_K},
+    {"q5_k", GGML_FTYPE_MOSTLY_Q5_K},
+    {"q6_k", GGML_FTYPE_MOSTLY_Q6_K},
 };
 
 void ggml_print_ftypes(FILE * fp) {
@@ -48,6 +53,11 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
         case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
         case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
+        case GGML_FTYPE_MOSTLY_Q2_K: qtype = GGML_TYPE_Q2_K; break;
+        case GGML_FTYPE_MOSTLY_Q3_K: qtype = GGML_TYPE_Q3_K; break;
+        case GGML_FTYPE_MOSTLY_Q4_K: qtype = GGML_TYPE_Q4_K; break;
+        case GGML_FTYPE_MOSTLY_Q5_K: qtype = GGML_TYPE_Q5_K; break;
+        case GGML_FTYPE_MOSTLY_Q6_K: qtype = GGML_TYPE_Q6_K; break;
         case GGML_FTYPE_UNKNOWN:
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
@@ -162,24 +172,17 @@ bool ggml_common_quantize_0(
 
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
-                    {
-                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q4_1:
-                    {
-                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q5_0:
-                    {
-                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q5_1:
-                    {
-                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
-                    } break;
                 case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements, hist_cur.data());
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
@@ -187,6 +190,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_I16:
                 case GGML_TYPE_I32:
                 case GGML_TYPE_Q8_1:
+                case GGML_TYPE_Q8_K:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/common-sdl.cpp

@@ -139,10 +139,13 @@ void audio_async::callback(uint8_t * stream, int len) {
         return;
     }
 
-    const size_t n_samples = len / sizeof(float);
+    size_t n_samples = len / sizeof(float);
 
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
+    if (n_samples > m_audio.size()) {
+        n_samples = m_audio.size();
+
+        stream += (len - (n_samples * sizeof(float)));
+    }
 
     //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
 
@@ -153,7 +156,7 @@ void audio_async::callback(uint8_t * stream, int len) {
             const size_t n0 = m_audio.size() - m_audio_pos;
 
             memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
+            memcpy(&m_audio[0], stream + n0 * sizeof(float), (n_samples - n0) * sizeof(float));
 
             m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
             m_audio_len = m_audio.size();

examples/common-sdl.h

@@ -41,7 +41,6 @@ private:
     std::mutex       m_mutex;
 
     std::vector<float> m_audio;
-    std::vector<float> m_audio_new;
     size_t             m_audio_pos = 0;
     size_t             m_audio_len = 0;
 };

examples/common.cpp

@@ -1,3 +1,5 @@
+#define _USE_MATH_DEFINES // for M_PI
+
 #include "common.h"
 
 // third-party utilities
@@ -6,43 +8,72 @@
 #include "dr_wav.h"
 
 #include <cmath>
+#include <cstring>
 #include <fstream>
 #include <regex>
+#include <locale>
+#include <codecvt>
+#include <sstream>
 
-#ifndef M_PI
-#define M_PI 3.14159265358979323846
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+// Function to check if the next argument exists
+std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+    if (i + 1 < argc && argv[i + 1][0] != '-') {
+        return argv[++i];
+    } else {
+        fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
+        gpt_print_usage(argc, argv, params);
+        exit(0);
+    }
+}
+
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
         if (arg == "-s" || arg == "--seed") {
-            params.seed = std::stoi(argv[++i]);
+            params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-t" || arg == "--threads") {
-            params.n_threads = std::stoi(argv[++i]);
+            params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-p" || arg == "--prompt") {
-            params.prompt = argv[++i];
+            params.prompt = get_next_arg(i, argc, argv, arg, params);
         } else if (arg == "-n" || arg == "--n_predict") {
-            params.n_predict = std::stoi(argv[++i]);
+            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-np" || arg == "--n_parallel") {
+            params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_k") {
-            params.top_k = std::stoi(argv[++i]);
+            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--top_p") {
-            params.top_p = std::stof(argv[++i]);
+            params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "--temp") {
-            params.temp = std::stof(argv[++i]);
+            params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-last-n") {
+            params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-penalty") {
+            params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-b" || arg == "--batch_size") {
-            params.n_batch = std::stoi(argv[++i]);
+            params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-c" || arg == "--context") {
+            params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--ignore-eos") {
+            params.ignore_eos = true;
         } else if (arg == "-m" || arg == "--model") {
-            params.model = argv[++i];
+            params.model = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-i" || arg == "--interactive") {
+            params.interactive = true;
+        } else if (arg == "-ip" || arg == "--interactive-port") {
+            params.interactive = true;
+            params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, params);
             exit(0);
         } else if (arg == "-f" || arg == "--file") {
-            if (++i > argc) {
-                fprintf(stderr, "Invalid file param");
-                break;
-            }
+            get_next_arg(i, argc, argv, arg, params);
             std::ifstream file(argv[i]);
             if (!file) {
                 fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
@@ -52,7 +83,10 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
-        } else {
+        } else if (arg == "-tt" || arg == "--token_test") {
+            params.token_test = get_next_arg(i, argc, argv, arg, params);
+        }
+        else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, params);
             exit(0);
@@ -73,11 +107,18 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "                        prompt to start generation with (default: random)\n");
     fprintf(stderr, "  -f FNAME, --file FNAME\n");
     fprintf(stderr, "                        load prompt from a file\n");
+    fprintf(stderr, "  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
+    fprintf(stderr, "                        test tokenization\n");
     fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
     fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
     fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
     fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
+    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
     fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "\n");
@@ -117,6 +158,10 @@ std::string replace(const std::string & s, const std::string & from, const std::
     return result;
 }
 
+void gpt_vocab::add_special_token(const std::string & token) {
+    special_tokens.push_back(token);
+}
+
 std::map<std::string, int32_t> json_parse(const std::string & fname) {
     std::map<std::string, int32_t> result;
 
@@ -208,8 +253,28 @@ std::map<std::string, int32_t> json_parse(const std::string & fname) {
     return result;
 }
 
-void gpt_vocab::add_special_token(const std::string & token) {
-    special_tokens.push_back(token);
+std::string convert_to_utf8(const std::wstring & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.to_bytes(input);
+}
+
+
+std::wstring convert_to_wstring(const std::string & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.from_bytes(input);
+}
+
+void gpt_split_words(std::string str, std::vector<std::string>& words) {
+    const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+    const std::regex re(pattern);
+    std::smatch m;
+
+    while (std::regex_search(str, m, re)) {
+        for (auto x : m) {
+            words.push_back(x);
+        }
+        str = m.suffix();
+    }
 }
 
 std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
@@ -218,62 +283,52 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     // first split the text into words
     {
         std::string str = text;
-        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
 
         // Generate the subpattern from the special_tokens vector if it's not empty
         if (!vocab.special_tokens.empty()) {
+            const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])");
             std::string special_tokens_subpattern;
             for (const auto & token : vocab.special_tokens) {
                 if (!special_tokens_subpattern.empty()) {
                     special_tokens_subpattern += "|";
                 }
-                special_tokens_subpattern += token;
+                special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)");
             }
 
-            // Modify the regex pattern with the generated special tokens subpattern
-            pat = special_tokens_subpattern + "|" + pat;
-        }
-
-        std::regex re(pat);
-        std::smatch m;
-
-        while (std::regex_search(str, m, re)) {
-            for (auto x : m) {
-                words.push_back(x);
+            std::regex re(special_tokens_subpattern);
+            std::smatch m;
+            // Split the text by special tokens.
+            while (std::regex_search(str, m, re)) {
+                // Split the substrings in-between special tokens into words.
+                gpt_split_words(m.prefix(), words);
+                // Add matched special tokens as words.
+                for (auto x : m) {
+                    words.push_back(x);
+                }
+                str = m.suffix();
             }
-            str = m.suffix();
+            // Remaining text without special tokens will be handled below.
         }
+
+        gpt_split_words(str, words);
     }
 
-    // find the longest tokens that form the words:
+    // find the longest token that forms each word in words:
     std::vector<gpt_vocab::id> tokens;
     for (const auto & word : words) {
-        if (word.size() == 0) continue;
-
-        int i = 0;
-        int n = word.size();
-        while (i < n) {
-            int j = n;
-            while (j > i) {
-                auto it = vocab.token_to_id.find(word.substr(i, j-i));
-                if (it != vocab.token_to_id.end()) {
+        for (int i = 0; i < (int) word.size(); ){
+            for (int j = word.size() - 1; j >= i; j--){
+                auto cand = word.substr(i, j-i+1);
+                auto it = vocab.token_to_id.find(cand);
+                if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab
                     tokens.push_back(it->second);
-                    i = j;
+                    i = j + 1;
                     break;
                 }
-                --j;
-            }
-            if (i == n) {
-                break;
-            }
-            if (j == i) {
-                auto sub = word.substr(i, 1);
-                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
-                    tokens.push_back(vocab.token_to_id.at(sub));
-                } else {
-                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
+                else if (j == i){ // word.substr(i, 1) has no matching
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
+                    i++;
                 }
-                ++i;
             }
         }
     }
@@ -281,6 +336,70 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
+std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+    std::vector<gpt_vocab::id> output;
+    std::stringstream ss(input);
+    std::string token;
+
+    while (std::getline(ss, token, delimiter)) {
+        output.push_back(std::stoi(token));
+    }
+
+    return output;
+}
+
+std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+    if (fpath_test.empty()){
+        fprintf(stderr, "%s : No test file found.\n", __func__);
+        return std::map<std::string, std::vector<gpt_vocab::id>>();
+    }
+
+    std::map<std::string, std::vector<gpt_vocab::id>> tests;
+
+    auto fin = std::ifstream(fpath_test, std::ios_base::in);
+    const char * delimeter = " => ";
+    const char del_tok = ',';
+    std::string line;
+    while (std::getline(fin, line)) {
+        size_t delimiterPos = line.find(delimeter);
+        if (delimiterPos != std::string::npos) {
+            std::string text = line.substr(0, delimiterPos);
+            std::string s_tokens = line.substr(delimiterPos + std::strlen(delimeter));
+            tests[text] = parse_tokens_from_string(s_tokens, del_tok);
+        }
+    }
+    return tests;
+}
+
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
+    std::map<std::string, std::vector<gpt_vocab::id>> tests = extract_tests_from_file(fpath_test);
+
+    size_t n_fails = 0;
+
+    for (const auto & test : tests) {
+        std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, test.first);
+
+        if (tokens != test.second){
+            n_fails++;
+
+            // print out failure cases
+            fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
+            fprintf(stderr, "%s : tokens in hf:   ", __func__);
+            for (const auto & t : test.second) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : tokens in ggml: ", __func__);
+            for (const auto & t : tokens) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+        }
+    }
+
+    fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
+}
+
 bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
     printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
 
@@ -380,6 +499,122 @@ gpt_vocab::id gpt_sample_top_k_top_p(
     return logits_id[idx].second;
 }
 
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng) {
+
+    int n_logits = vocab.id_to_token.size();
+
+    const auto * plogits = logits;
+
+    const auto last_n_tokens = std::vector<int32_t>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_data_size);
+
+    if (temp <= 0) {
+        // select the token with the highest logit directly
+        float max_logit = plogits[0];
+        gpt_vocab::id max_id = 0;
+
+        for (int i = 1; i < n_logits; ++i) {
+            if (plogits[i] > max_logit) {
+                max_logit = plogits[i];
+                max_id = i;
+            }
+        }
+        return max_id;
+    }
+
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const float scale = 1.0f/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
+            // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
+            if (repeat_last_n > 0 && std::find(last_n_tokens.end()-repeat_last_n, last_n_tokens.end(), i) != last_n_tokens.end()) {
+                // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if (plogits[i] < 0.0f) {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
+                } else {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
+                }
+            } else {
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
+            }
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+//    printf("\n");
+//    for (int i = 0; i < (int) probs.size(); i++) {
+//    for (int i = 0; i < 10; i++) {
+//        printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+//    }
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+
+}
+
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
     std::vector<uint8_t> wav_data; // used for pipe input from stdin
@@ -537,3 +772,46 @@ float similarity(const std::string & s0, const std::string & s1) {
 
     return 1.0f - (dist / std::max(s0.size(), s1.size()));
 }
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-i" || arg == "--inp") {
+            params.fname_inp = argv[++i];
+        } else if (arg == "-o" || arg == "--out") {
+            params.fname_out = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
+    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
+    fprintf(stderr, "  -o FNAME, --out FNAME\n");
+    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}

examples/common.h

@@ -7,27 +7,39 @@
 #include <vector>
 #include <random>
 #include <thread>
+#include <ctime>
+#include <fstream>
 
 #define COMMON_SAMPLE_RATE 16000
 
 //
-// CLI argument parsing
+// GPT CLI argument parsing
 //
 
 struct gpt_params {
-    int32_t seed      = -1; // RNG seed
-    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_predict = 200; // new tokens to predict
+    int32_t seed         = -1;   // RNG seed
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict    = 200;  // new tokens to predict
+    int32_t n_parallel   = 1;    // number of parallel streams
+    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
+    int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
+
+    bool ignore_eos = false; // ignore EOS token when generating text
 
     // sampling parameters
-    int32_t top_k = 40;
-    float   top_p = 0.9f;
-    float   temp  = 0.9f;
+    int32_t top_k          = 40;
+    float   top_p          = 0.9f;
+    float   temp           = 0.9f;
+    int32_t repeat_last_n  = 64;
+    float   repeat_penalty = 1.00f;
 
-    int32_t n_batch = 8; // batch size for prompt processing
+    std::string model      = "models/gpt-2-117M/ggml-model.bin"; // model path
+    std::string prompt     = "";
+    std::string token_test = "";
 
-    std::string model = "models/gpt-2-117M/ggml-model.bin"; // model path
-    std::string prompt;
+    bool    interactive      = false;
+    int32_t interactive_port = -1;
 };
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@@ -61,6 +73,12 @@ struct gpt_vocab {
 // poor-man's JSON parsing
 std::map<std::string, int32_t> json_parse(const std::string & fname);
 
+std::string convert_to_utf8(const std::wstring & input);
+
+std::wstring convert_to_wstring(const std::string & input);
+
+void gpt_split_words(std::string str, std::vector<std::string>& words);
+
 // split text into tokens
 //
 // ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
@@ -73,6 +91,15 @@ std::map<std::string, int32_t> json_parse(const std::string & fname);
 //
 std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
 
+// test outputs of gpt_tokenize
+//
+//   - compare with tokens generated by the huggingface tokenizer
+//   - test cases are chosen based on the model's main language (under 'prompt' directory)
+//   - if all sentences are tokenized identically, print 'All tests passed.'
+//   - otherwise, print sentence, huggingface tokens, ggml tokens
+//
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test);
+
 // load the tokens from encoder.json
 bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
 
@@ -92,6 +119,18 @@ gpt_vocab::id gpt_sample_top_k_top_p(
         double temp,
         std::mt19937 & rng);
 
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng);
+
 //
 // Audio utils
 //
@@ -105,6 +144,104 @@ bool read_wav(
         std::vector<std::vector<float>> & pcmf32s,
         bool stereo);
 
+// Write PCM data into WAV audio file
+class wav_writer {
+private:
+    std::ofstream file;
+    uint32_t dataSize = 0;
+    std::string wav_filename;
+
+    bool write_header(const uint32_t sample_rate,
+                      const uint16_t bits_per_sample,
+                      const uint16_t channels) {
+
+        file.write("RIFF", 4);
+        file.write("\0\0\0\0", 4);    // Placeholder for file size
+        file.write("WAVE", 4);
+        file.write("fmt ", 4);
+
+        const uint32_t sub_chunk_size = 16;
+        const uint16_t audio_format = 1;      // PCM format
+        const uint32_t byte_rate = sample_rate * channels * bits_per_sample / 8;
+        const uint16_t block_align = channels * bits_per_sample / 8;
+
+        file.write(reinterpret_cast<const char *>(&sub_chunk_size), 4);
+        file.write(reinterpret_cast<const char *>(&audio_format), 2);
+        file.write(reinterpret_cast<const char *>(&channels), 2);
+        file.write(reinterpret_cast<const char *>(&sample_rate), 4);
+        file.write(reinterpret_cast<const char *>(&byte_rate), 4);
+        file.write(reinterpret_cast<const char *>(&block_align), 2);
+        file.write(reinterpret_cast<const char *>(&bits_per_sample), 2);
+        file.write("data", 4);
+        file.write("\0\0\0\0", 4);    // Placeholder for data size
+
+        return true;
+    }
+
+    // It is assumed that PCM data is normalized to a range from -1 to 1
+    bool write_audio(const float * data, size_t length) {
+        for (size_t i = 0; i < length; ++i) {
+            const int16_t intSample = data[i] * 32767;
+            file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
+            dataSize += sizeof(int16_t);
+        }
+        if (file.is_open()) {
+            file.seekp(4, std::ios::beg);
+            uint32_t fileSize = 36 + dataSize;
+            file.write(reinterpret_cast<char *>(&fileSize), 4);
+            file.seekp(40, std::ios::beg);
+            file.write(reinterpret_cast<char *>(&dataSize), 4);
+            file.seekp(0, std::ios::end);
+        }
+        return true;
+    }
+
+    bool open_wav(const std::string & filename) {
+        if (filename != wav_filename) {
+            if (file.is_open()) {
+                file.close();
+            }
+        }
+        if (!file.is_open()) {
+            file.open(filename, std::ios::binary);
+            wav_filename = filename;
+            dataSize = 0;
+        }
+        return file.is_open();
+    }
+
+public:
+    bool open(const std::string & filename,
+              const    uint32_t   sample_rate,
+              const    uint16_t   bits_per_sample,
+              const    uint16_t   channels) {
+
+        if (open_wav(filename)) {
+            write_header(sample_rate, bits_per_sample, channels);
+        } else {
+            return false;
+        }
+
+        return true;
+    }
+
+    bool close() {
+        file.close();
+        return true;
+    }
+
+    bool write(const float * data, size_t length) {
+        return write_audio(data, length);
+    }
+
+    ~wav_writer() {
+        if (file.is_open()) {
+            file.close();
+        }
+    }
+};
+
+
 // Apply a high-pass frequency filter to PCM audio
 // Suppresses frequencies below cutoff Hz
 void high_pass_filter(
@@ -123,3 +260,20 @@ bool vad_simple(
 
 // compute similarity between two strings using Levenshtein distance
 float similarity(const std::string & s0, const std::string & s1);
+
+//
+// SAM argument parsing
+//
+
+struct sam_params {
+    int32_t seed      = -1; // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    std::string model     = "models/sam-vit-b/ggml-model-f16.bin"; // model path
+    std::string fname_inp = "img.jpg";
+    std::string fname_out = "img.out";
+};
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params);
+
+void sam_print_usage(int argc, char ** argv, const sam_params & params);

examples/grammar-parser.cpp

@@ -0,0 +1,423 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from whisper.cpp
+    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<whisper_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<whisper_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({WHISPER_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum whisper_gretype start_type = WHISPER_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = WHISPER_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum whisper_gretype type = last_sym_start < out_elements.size()
+                        ? WHISPER_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({WHISPER_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<whisper_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({WHISPER_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({WHISPER_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<whisper_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({WHISPER_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({WHISPER_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    bool is_char_element(whisper_grammar_element elem) {
+        switch (elem.type) {
+            case WHISPER_GRETYPE_CHAR:           return true;
+            case WHISPER_GRETYPE_CHAR_NOT:       return true;
+            case WHISPER_GRETYPE_CHAR_ALT:       return true;
+            case WHISPER_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
+                case WHISPER_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case WHISPER_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case WHISPER_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case WHISPER_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case WHISPER_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                case WHISPER_GRETYPE_ALT:
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                case WHISPER_GRETYPE_CHAR_NOT:
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<whisper_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != WHISPER_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with WHISPER_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            whisper_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case WHISPER_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case WHISPER_GRETYPE_CHAR_ALT:
+                    case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
+        std::vector<const whisper_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

examples/grammar-parser.h

@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by whisper.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "whisper.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                   symbol_ids;
+        std::vector<std::vector<whisper_grammar_element>> rules;
+
+        std::vector<const whisper_grammar_element *>      c_rules() const;
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}

examples/livestream.sh

@@ -48,7 +48,7 @@ if [ -n "$3" ]; then
 fi
 
 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
 
 # list available models
 function list_models {

examples/lsp/CMakeLists.txt

@@ -0,0 +1,9 @@
+if (WHISPER_SDL2)
+    # stream
+    set(TARGET lsp)
+    add_executable(${TARGET} lsp.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/lsp/README.md

@@ -0,0 +1,104 @@
+# Language Server
+
+This example consists of a simple language server to expose both unguided
+and guided (command) transcriptions by sending json messages over stdout/stdin
+as well as a rather robust vim plugin that makes use of the language server.
+
+## Vim plugin quick start
+
+Compile the language server with
+
+```bash
+make lsp
+```
+Install the plugin itself by copying or symlinking whisper.vim into ~/.vim/autoload/
+
+In your vimrc, set the path of your whisper.cpp directory and optionally add some keybinds.
+
+```vim
+let g:whisper_dir = "~/whisper.cpp"
+" Start listening for commands when Ctrl - g is pressed in normal mode
+nnoremap <C-G> call whisper#requestCommands()<CR>
+" Start unguided transcription when Ctrl - g is pressed in insert mode
+inoremap <C-G> <Cmd>call whisper#doTranscription()<CR>
+```
+
+## Vim plugin usage
+
+The vim plugin was designed to closely follow the mnemonics of vim
+
+`s:spoken_dict` is used to translate keys to their spoken form.
+
+
+Keys corresponding to a string use that spoken value normally and when a motion is expected, but use the key itself when a character is expected.  
+Keys corresponding to a dict, like `i`, can have manual difinitions given to each possible commandset.
+
+0 is normal (insert), 1 is motion (inside), 2 is it's usage as a single key ([till] i), and 3 is it's usage in an area selection (s -> [around] sentence)
+
+Some punctuation items, like `-` are explicitly given pronunciations to prevent them from being picked as punctuation instead of an actual command word.
+
+Not all commands will tokenize to a single token and this can interfere with interpretation. "yank" as an example, takes multiple tokens and correspondingly, will give more accurate detection when only the first "ya" is used. While it could be changed to something else that is a single token (copy), value was placed on maintaining vim mnemonics.
+
+Commands that would normally move the editor into insert mode (insert, append, open, change) will begin unguided transcription.
+Unguided transcription will end when a speech segment ends in exit.
+Presence of punctuation can be designated by whether or not you add a pause between the previous speech segment and exit.
+Exiting only occurs if exit is the last word, so "Take the first exit on your right" would not cause transcription to end.
+
+After a command is evaluated, the plugin will continue listening for the next command.
+
+While in command mode, "Exit" will end listening.
+
+A best effort approach is taken to keep track of audio that is recorded while a previous chunk is still processing and immediately interpret it afterwards, but the current voice detection still needs a fairly sizable gap to determine when a command has been spoken.
+
+Log information is sent to a special `whisper_log` buffer and can be accessed with
+```vim
+:e whisper_log
+```
+
+## Vim plugin configuration
+
+`g:whisper_dir`  
+A full path to the whisper.cpp repo. It can be expanded in the definition like so:
+```vim
+let g:whisper_dir = expand("~/whisper.cpp/")
+```
+(The WHISPER_CPP_HOME environment variable is also checked for users of the existing whisper.nvim script)
+
+`g:whisper_lsp_path`  
+Can be used to manually set the path to the language server.
+If not defined, it will be inferred from the above whisper_dir
+
+`g:whisper_model_path`  
+A full path to the model to load. If not defined, it will default to ggml-base.en.bin
+
+`g:whisper_user_commands`  
+A dictionary of spoken commands that correspond to either strings or funcrefs.
+This can be used to create connections with other user plugins, for example
+```vim
+let g:whisper_user_commands = {"gen": "llama#doLlamaGen"}
+```
+will trigger the llama.cpp plugin to begin generation when "gen" is spoken
+
+## Language server methods
+
+`registerCommandset`  
+`params` is a list of strings that should be checked for with this commandset. The server prepends a space to these strings before tokenizing.  
+Responds with  
+`result.index` an integer index for the commandset registered, which should be included when initiating a guided transcription to select this commandset.
+Will return an error if any of the commands in the commandset have duplicate tokenizations
+
+`guided`  
+`params.commandset_index` An index returned by a corresponding commandset registration. If not set, the most recently registered commandset is used.
+`params.timestamp` A positive unsigned integer which designates a point in time which audio should begin processing from. If left blank, the start point of audio processing will be the moment the message is recieved. This should be left blank unless you have a timestamp from a previous response.  
+Responds with  
+`result.command_index` The numerical index (starting from 0) of the detected command in the selected commandset
+`result.command_text` A string containing the command as provided in the commandset
+`result.timestamp` A positive unsigned integer that designates the point in time which audio stopped being processed at. Pass this timestamp back in a subsequent message to mask the latency of transcription.
+
+`unguided`  
+`params.no_context` Sets the corresponding whisper `no_context` param. Defaults to true. Might provide more accurate results for consecutive unguided transcriptions if those after the first are set to false.
+`params.prompt` If provided, sets the initial prompt used during transcription.
+`params.timestamp` A positive unsigned integer which designates a point in time which audio should begin processing from. If left blank, the start point of audio processing will be the moment the message is recieved. This should be left blank unless you have a timestamp from a previous response.  
+Responds with  
+`result.transcription` A string containing the transcribed text.  N.B. This will almost always start with a space due to how text is tokenized.
+`result.timestamp` A positive unsigned integer that designates the point in time which audio stopped being processed at. Pass this timestamp back in a subsequent message to mask the latency of transcription.

examples/lsp/lsp.cpp

@@ -0,0 +1,463 @@
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "json.hpp"
+
+#include <iostream>
+#include <cassert>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+#include <deque>
+#include <set>
+
+using json = nlohmann::json;
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t prompt_ms  = 5000;
+    int32_t command_ms = 8000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool use_gpu       = true;
+
+    std::string language  = "en";
+    std::string model     = "models/ggml-base.en.bin";
+};
+struct command {
+    std::vector<whisper_token> tokens;
+    std::string plaintext;
+};
+struct commandset {
+    std::vector<struct command> commands;
+    std::vector<whisper_token> prompt_tokens;
+    // TODO: Store longest command?
+    // Multi-token commands should have probabilities of subsequent logits
+    // given that the prior logit is correct.
+    // In this case, all commands must be iterated.
+    // This however, is likely highly involved as different tokens
+    // almost certainly have different spoken lengths
+    // It would also have performance implications equivalent to a beam search
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms" || arg == "--command-ms")    { params.command_ms    = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,         --help           [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,       --threads N      [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -pms N,     --prompt-ms N    [%-7d] prompt duration in milliseconds\n",             params.prompt_ms);
+    fprintf(stderr, "  -cms N,     --command-ms N   [%-7d] command duration in milliseconds\n",            params.command_ms);
+    fprintf(stderr, "  -c ID,      --capture ID     [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,      --max-tokens N   [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "\n");
+}
+uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
+    using namespace std::chrono;
+    uint64_t time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+    uint64_t start_time = time_now;
+    if (jparams.contains("timestamp")) {
+        start_time = jparams.at("timestamp");
+    }
+    if(time_now - start_time < 500) {
+        //wait for a backlog of audio
+        std::this_thread::sleep_for(milliseconds(500 - (time_now - start_time)));
+        time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+    } else if (time_now - start_time > 1000) {
+        audio.get(time_now-start_time, pcmf32);
+        size_t max_offset = pcmf32.size() - WHISPER_SAMPLE_RATE;
+        for(size_t offset=0;offset < max_offset;offset+=WHISPER_SAMPLE_RATE/10) {
+            std::vector<float> audio_chunk(&pcmf32[offset], &pcmf32[offset+WHISPER_SAMPLE_RATE]);
+            if(::vad_simple(audio_chunk, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                pcmf32.resize(offset+WHISPER_SAMPLE_RATE);
+                if (offset*1000/WHISPER_SAMPLE_RATE+1000 > maxlength_ms) {
+                    //remove samples from the beginning
+                    pcmf32.erase(pcmf32.begin(),pcmf32.end()-(maxlength_ms*WHISPER_SAMPLE_RATE/1000));
+                    fprintf(stderr, "Shortened samples");
+                }
+                return start_time + offset*1000/WHISPER_SAMPLE_RATE+1000;
+            }
+        }
+    }
+    size_t window_duration = std::max((uint64_t)1000, time_now-start_time);
+    audio.get(window_duration, pcmf32);
+    while (!::vad_simple(pcmf32, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+        std::this_thread::sleep_for(milliseconds(100));
+        time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
+        window_duration = std::max((uint64_t)1000,time_now-start_time);
+        audio.get(window_duration, pcmf32);
+    }
+    if (time_now - start_time > maxlength_ms) {
+        audio.get(maxlength_ms, pcmf32);
+    } else {
+        audio.get(time_now - start_time, pcmf32);
+    }
+
+    return time_now;
+}
+
+json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
+    std::vector<whisper_token> prompt_tokens;
+    std::vector<float> pcmf32;
+    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 10000U, pcmf32);
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    if (jparams.contains("prompt")) {
+        // unlikely to see much use. Under normal circumstances, no_context would be set to false
+        std::string prompt = jparams.at("prompt");
+        prompt_tokens.resize(1024);
+        int n = whisper_tokenize(ctx, prompt.c_str(), prompt_tokens.data(), 1024);
+        prompt_tokens.resize(n);
+
+        wparams.prompt_tokens    = prompt_tokens.data();
+        wparams.prompt_n_tokens  = prompt_tokens.size();
+    }
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = false;
+    wparams.translate        = params.translate;
+    wparams.no_context       = jparams.value("no_context", true);
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+    wparams.suppress_non_speech_tokens = true;
+    // run the transformer and a single decoding pass
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+        throw json{
+            {"code", -32803},
+            {"message", "ERROR: whisper_full() failed"}
+        };
+    }
+    std::string result = whisper_full_get_segment_text(ctx,0);
+    return json {
+        {"transcription", result},
+        {"timestamp", unprocessed_audio_timestamp}
+    };
+}
+
+// command-list mode
+// guide the transcription to match the most likely command from a provided list
+json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
+    struct commandset cs = commandset_list[jparams.value("commandset_index", commandset_list.size()-1)];
+    std::vector<float> pcmf32;
+    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 2000U, pcmf32);
+
+    fprintf(stderr, "%s: Speech detected! Processing ...\n", __func__);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = false;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = 1;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    // TODO: Do some time testing. Does an overly long prompt slow down processing?
+    // Set up command sets/precompute prompts
+    wparams.prompt_tokens    = cs.prompt_tokens.data();
+    wparams.prompt_n_tokens  = cs.prompt_tokens.size();
+    // TODO: properly expose as option
+    wparams.suppress_non_speech_tokens = true;
+
+    // run the transformer and a single decoding pass
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+        throw json{
+            {"code", -32803},
+            {"message", "ERROR: whisper_full() failed"}//TODO: format string (sprintf?)
+        };
+    }
+
+    // estimate command probability
+    // NOTE: not optimal
+    {
+        const auto * logits = whisper_get_logits(ctx);
+
+        std::vector<float> probs(whisper_n_vocab(ctx), 0.0f);
+
+        // compute probs from logits via softmax
+        {
+            float max = -1e9;
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                max = std::max(max, logits[i]);
+            }
+
+            float sum = 0.0f;
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                probs[i] = expf(logits[i] - max);
+                sum += probs[i];
+            }
+
+            for (int i = 0; i < (int) probs.size(); ++i) {
+                probs[i] /= sum;
+            }
+        }
+
+        std::vector<std::pair<float, int>> probs_id;
+
+        // In my testing, the most verbose token is always the desired.
+        // TODO: Trim commandset struct once efficacy has been verified
+        for (int i = 0; i < (int) cs.commands.size(); ++i) {
+            probs_id.emplace_back(probs[cs.commands[i].tokens[0]], i);
+        }
+
+        // sort descending
+        {
+            using pair_type = decltype(probs_id)::value_type;
+            std::sort(probs_id.begin(), probs_id.end(), [](const pair_type & a, const pair_type & b) {
+                    return a.first > b.first;
+                    });
+        }
+        int id = probs_id[0].second;
+        return json{
+            {"command_index", id},
+                {"command_text", cs.commands[id].plaintext},
+                {"timestamp", unprocessed_audio_timestamp},
+        };
+    }
+}
+
+json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
+    // TODO: check for token collision
+    struct commandset cs;
+
+    std::string  k_prompt = " select one from the available words: ";
+    std::set<whisper_token> token_set;
+    whisper_token tokens[32];
+    for (std::string s : jparams) {
+        std::vector<whisper_token> token_vec;
+        // The existing command implementation uses a nested for loop to tokenize single characters
+        // I fail to see the purpose of this when ' a' has a wholly different pronunciation than the start of ' apple'
+        const int n = whisper_tokenize(ctx, (" " + s).c_str(), tokens, 32);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, s.c_str());
+            return 3;
+        }
+        token_vec.push_back(tokens[0]);
+        if (!token_set.insert(tokens[0]).second) {
+            fprintf(stderr, "%s: warning: %s is a duplicate of an existing token\n", __func__, s.c_str());
+            throw json{
+                {"code",-31000},
+                {"message", "Duplicate token in token set: " + s}
+            };
+        }
+        if (n > 1) {// empty string if n=0? Should never occur
+            fprintf(stderr, "%s: error: command is more than a single token: %s\n", __func__, s.c_str());
+        }
+        struct command command = {token_vec, s};
+        cs.commands.push_back(command);
+        k_prompt += s;
+    }
+    k_prompt = k_prompt.substr(0,k_prompt.length()-2) + ". Selected word:";
+    cs.prompt_tokens.resize(1024);
+    int n = whisper_tokenize(ctx, k_prompt.c_str(), cs.prompt_tokens.data(), 1024);
+    cs.prompt_tokens.resize(n);
+    // prepare response
+    int index = commandset_list.size();
+    commandset_list.push_back(cs);
+    return json{{"index",index}};
+}
+json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
+    // whisper_state has the pertinent offsets, but there also seem to be a large
+    // number of scratch buffers that would prevent rewinding context in a manner similar to llama
+    // I'll give this a another pass once everything else is implemented,
+    // but for now, it's unsupported
+    throw json {
+        {"code", -32601},
+            {"message", "Seeking is not yet supported."}
+    };
+}
+json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
+    // See: https://www.jsonrpc.org/specification
+    json id = body.at("id");
+    try {
+        std::string version = body.at("jsonrpc");
+        if (version != "2.0") {
+            // unsupported version
+            throw json{
+                {"code", -3260},
+                {"message", "invalid jsonrpc version"}
+            };
+        }
+        std::string method = body.at("method");
+        json jparams = json{{"dummy", "dummy"}};
+        if (body.contains("params"))
+            jparams = body.at("params");
+        json res;
+        // TODO: be consistent about argument order
+        fprintf(stderr, "Dispatching a job\n");
+        if (method == "unguided")                { res = unguided_transcription(ctx, audio, jparams, params); }
+        else if (method == "guided")             { res = guided_transcription(ctx, audio, params, jparams, commandset_list); }
+        else if (method == "seek")               { res = seek(ctx, audio, jparams); }
+        else if (method == "registerCommandset") { res = register_commandset(ctx, jparams, commandset_list); }
+        else if (method == "echo")               { res = jparams; }
+
+
+        return json{
+            {"jsonrpc", "2.0"},
+                {"result", res},
+                {"id", id}
+        };
+    } catch(json ex) {
+        return json {
+            {"jsonrpc", "2.0"},
+                {"error", ex},
+                {"id", id}
+        };
+    }
+}
+
+void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    std::deque<json> jobqueue;
+    std::vector<struct commandset> commandset_list;
+    while (true) {
+        // For eventual cancellation support, shouldn't block if job exists
+        if (std::cin.rdbuf()->in_avail() > 22 || jobqueue.size() == 0) {
+            int content_length;
+            if (scanf("Content-Length: %d", &content_length) != 1) {
+                fprintf(stderr, "Could not read input: %d", std::cin.peek());
+                return;
+            }
+            // scanf leaves the new lines intact
+            std::cin.ignore(2);
+            if (std::cin.peek() != 13) {
+                // Content-Type. jsonrpc necessitates utf8.
+                std::cin.ignore(200,10);
+            }
+            std::cin.ignore(2);
+            // A message is being sent and blocking is acceptable
+            std::string content(content_length,'\0');
+            std::cin.read(&content[0], content_length);
+            json job = json::parse(content);
+            // TODO: Some messages(cancellation) should skip queue here
+            if (job.is_array()) {
+                // response must also be batched. Will implement later
+                // for (subjob : job.begin())
+                // TODO: At the very least respond with an unsupported error.
+            } else {
+                jobqueue.push_back(job);
+            }
+        }
+        assert(jobqueue.size() > 0);
+        json job = jobqueue.front();
+        json resp = parse_job(job, ctx, audio, params, commandset_list);
+        if (resp != "unfinished") {
+            jobqueue.pop_front();
+            // send response
+            std::string data = resp.dump(-1, ' ', false, json::error_handler_t::replace);
+            fprintf(stdout, "Content-Length: %d\r\n\r\n%s\n", (int)data.length()+1, data.c_str());
+            std::cout.flush();
+
+        }
+    }
+}
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+    // TODO: Investigate why this is required. An extra second of startup latency is not great
+    // wait for 1 second to avoid any buffered noise
+    std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+    audio.clear();
+    // TODO: consider some sort of indicator to designate loading has finished?
+    // Potentially better for the client to just start with a non-blocking message (register commands)
+    process_loop(ctx, audio, params);
+
+    audio.pause();
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}

examples/lsp/whisper.vim

@@ -0,0 +1,362 @@
+if !exists("g:whisper_dir")
+    let g:whisper_dir = expand($WHISPER_CPP_HOME)
+    if g:whisper_dir == ""
+        echoerr "Please provide a path to the whisper.cpp repo in either the $WHISPER_CPP_HOME environment variable, or g:whisper_dir"
+    endif
+endif
+if !exists("g:whisper_lsp_path")
+    let g:whisper_lsp_path = g:whisper_dir .. "lsp"
+    if !filereadable(g:whisper_lsp_path)
+        echoerr "Was not able to locate a lsp executable at: " .. g:whisper_lsp_path
+        throw "Executable not found"
+    endif
+endif
+if !exists("g:whisper_model_path")
+    " TODO: allow custom paths relative to the repo dir
+    let g:whisper_model_path = g:whisper_dir .. "models/ggml-base.en.bin"
+    if !filereadable(g:whisper_model_path)
+        echoerr "Could not find model at: " .. g:whisper_model_path
+        throw "Model not found"
+    endif
+endif
+let s:output_buffer = bufnr("whisper_log", v:true)
+call setbufvar(s:output_buffer,"&buftype","nofile")
+let s:lsp_command = [g:whisper_lsp_path,"-m",g:whisper_model_path]
+" For faster execution. TODO: server load multiple models/run multiple servers?
+" let s:lsp_command = [g:whisper_lsp_path, "-m", g:whisper_dir .. "models/ggml-tiny.en.bin", "-ac", "128"]
+
+" requestCommands([params_dict])
+func whisper#requestCommands(...)
+    let l:req = {"method": "guided", "params": {"commandset_index": 0}}
+    if a:0 > 0
+        call extend(l:req.params, a:1)
+    endif
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:commandCallback", [l:req.params, 0])})
+endfunction
+
+" doTranscription([params_dict])
+func whisper#doTranscription(...)
+    let l:req = {"method": "unguided", "params": {}}
+    if a:0 > 0
+        call extend(l:req.params, a:1)
+    endif
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:transcriptionCallback", [function("s:insertText"),function("s:endTranscription")])})
+endfunction
+
+" For testing
+func whisper#uppertest(cha)
+    echo tr(a:cha, s:c_lowerkeys, s:c_upperkeys)
+endfunction
+
+
+" (upper, exit, count, motion, command, insert/append, save run) "base"
+" (upper, exit, count, motion, command, inside/around)           "motion/visual"
+" (upper, exit, count, motion, line,    inside/around)           "command already entered"
+" (upper, exit, key,                                 )           "from/till"
+
+" upper and lower keys is used to translate between cases with tr
+" Must be sunchronized
+let s:c_lowerkeys = "1234567890-=qwertyuiop[]\\asdfghjkl;'zxcvbnm,./\""
+let s:c_upperkeys = "!@#$%^&*()_+QWERTYUIOP{}|ASDFGHJKL:\"ZXCVBNM<>?'"
+let s:c_count = split("1234567890\"",'\zs')
+let s:c_command = split("ryuogpdxcv.iam", '\zs')
+let s:c_motion = split("wetf'hjklnb$^)",'\zs')
+" object words: Word, Sentence, Paragraph, [, (, <, Tag, {. ", '
+let s:c_area = split("wsp])>t}\"'",'\zs')
+"Special commands.
+let s:c_special_always = ["exit", "upper"]
+let s:c_special_normal = ["save", "run", "space"]
+
+" If not in dict, key is spoken word,
+" If key resolves to string, value is used for normal/motion, but key for chars
+" If key resolves to dict, {0: "normal",1: "motion",2:"single char",3: "area"}
+" Missing entries fall back as follows {0: "required", 1: 0, 2: "key", 3: 0}
+let s:spoken_dict = {"w": "word", "e": "end", "r": "replace", "t": {0: "till", 3: "tag"}, "y": "yank", "u": "undo", "i": {0: "insert", 1: "inside"}, "o": "open", "p": {0: "paste", 3: "paragraph"},  "a": {0: "append", 1: "around"}, "s": {0: "substitute", 3: "sentence"}, "d": "delete", "f": "from", "g": "go", "h": "left", "j": "down", "k": "up", "l": "right", "c": "change", "v": "visual", "b": "back", "n": "next", "m": "mark", ".": {0: "repeat", 2: "period"}, "]": {0: "bracket", 2: "bracket"}, "'": {0: "jump", 2: "apostrophe", 3:  "apostrophe"}, '"': {0: 'register', 2: "quotation", 3: "quotation"}, "-": {0: "minus", 2: "minus"}, "$": {0: "dollar", 2: "dollar"}, "^": {0: "carrot", 2: "carrot"}, ")": {0: "sentence", 2: "parenthesis",  3: "parenthesis"}, "}": {0: "paragraph", 2: "brace", 3: "brace"}, ">": {0: "indent", 2: "angle", 3: "angle"}}
+
+" Give this another pass. This seems overly hacky even if it's functional
+let s:sub_tran_msg = ""
+func s:subTranProg(msg)
+    if s:sub_tran_msg != ""
+        let s:sub_tran_msg = s:sub_tran_msg .. a:msg
+        if mode() !=? 'v'
+            exe "normal" "u" .. s:sub_tran_msg
+        endif
+    else
+        if s:command_backlog == ""
+            " this should not occur
+            call s:logCallback(0, "Warning: Encountered sub transcription without prior command")
+            let s:command_backlog = "a"
+        endif
+        if a:msg[0] == ' '
+            let s:sub_tran_msg = s:command_backlog .. a:msg[1:-1]
+        else
+            let s:sub_tran_msg = s:command_backlog  .. a:msg
+        endif
+        if mode() !=? 'v'
+            exe "normal" s:sub_tran_msg
+        endif
+    endif
+    call appendbufline(s:output_buffer, "$", s:sub_tran_msg ..  ":" .. string(a:msg ))
+endfunction
+
+func s:subTranFinish(params, timestamp)
+    let s:repeat_command = s:sub_tran_msg
+    " Visual selection is lot if used with streaming, so streaming of partial
+    " transcriptions is disabled in visual mode
+    if mode() ==? 'v'
+        exe "normal" s:sub_tran_msg
+    endif
+    let s:sub_tran_msg = ""
+    let s:command_backlog = ""
+    exe "normal a\<C-G>u"
+    let l:params = a:params
+    let l:params.timestamp = a:timestamp
+    if exists("l:params.commandset_index")
+        unlet l:params.commandset_index
+    endif
+    call whisper#requestCommands(a:params)
+endfunction
+
+func s:logCallback(channel, msg)
+    call appendbufline(s:output_buffer,"$",a:msg)
+endfunction
+
+
+func s:transcriptionCallback(progressCallback, finishedCallback, channel, msg)
+    let l:tr = a:msg.result.transcription
+
+    let l:ex_ind = match(tolower(l:tr),"exit", len(l:tr)-6)
+    " The worst case I've observed so far is " Exit.", which is 6 characters
+    if l:ex_ind != -1
+        call a:progressCallback(strpart(l:tr,0,l:ex_ind-1))
+        call a:finishedCallback(a:msg.result.timestamp)
+    else
+        call a:progressCallback(l:tr)
+        let req = {"method": "unguided", "params": {"timestamp": a:msg.result.timestamp, "no_context": v:true}}
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [a:progressCallback, a:finishedCallback])})
+    endif
+endfunc
+func s:insertText(msg)
+    exe "normal a" .. a:msg
+endfunction
+func s:endTranscription(timestamp)
+    call appendbufline(s:output_buffer, "$", "Ending unguided transcription")
+endfunction
+
+
+
+" If a command does not include a whole actionable step, attempting to execute
+" it discards the remainder of things. There is likely a simpler solution,
+" but it can be made functional now by storing a backbuffer until actionable
+let s:command_backlog = ""
+let s:repeat_command = ""
+let s:preceeding_upper = v:false
+func s:commandCallback(params, commandset_index, channel, msg)
+    let l:command_index = a:msg.result.command_index
+    let l:do_execute = v:false
+    let l:next_mode = a:commandset_index
+    let l:command = s:commandset_list[a:commandset_index][l:command_index]
+    call s:logCallback(0, string(a:msg) .. " " .. a:commandset_index .. " " .. l:command)
+    if l:command_index == 0
+        "exit
+        "if s:command_backlog == ""
+        call s:logCallback(0,"Stopping command mode")
+        echo "No longer listening"
+        let s:command_backlog = ""
+        return
+        "else
+        " Legacy code to clear an existing buffer with exit.
+        " Was found to be rarely desired and is better introduced as a
+        " standalone command (clear?)
+        "   call s:logCallback(0,"Clearing command_backlog" .. s:command_backlog)
+        "   let s:command_backlog = ""
+        "   let s:preceeding_upper = v:false
+        " endif
+    elseif l:command_index == 1
+        " upper
+        let s:preceeding_upper = !s:preceeding_upper
+    elseif l:command == "save"
+        " save and run can only happen in commandset 0,
+        exe "w"
+    elseif l:command == "run"
+        exe "make run"
+    elseif l:command == "space"
+        exe "normal i \<ESC>l"
+    elseif has_key(s:c_user, l:command)
+        let Userfunc = s:c_user[l:command]
+        if type(Userfunc) == v:t_string
+            let Userfunc = function(Userfunc)
+        endif
+        call Userfunc()
+    else
+        if s:preceeding_upper
+            " Upper should keep commandset
+            let s:preceeding_upper = v:false
+            let l:visual_command = tr(l:command, s:c_lowerkeys, s:c_upperkeys)
+        else
+            let l:visual_command = l:command
+        endif
+        echo s:command_backlog .. " - " .. l:visual_command
+        let s:command_backlog = s:command_backlog .. l:visual_command
+        if a:commandset_index == 2 || a:commandset_index == 3
+            " single key, either completes motion, replace, or register
+            " Should move to execute unless part of a register
+            " Change will be caught at execute
+            if s:command_backlog[-2:-2] !=# '"'
+                call s:logCallback(0,"not register")
+                let l:do_execute = v:true
+            end
+            let l:next_mode = 0
+            " commandset index only matters for a/i
+        elseif (l:command == "a" || l:command == "i") && a:commandset_index == 1
+            " inside/around. Is commandset 3
+            let l:next_mode = 3
+        elseif l:command ==# '"'
+            let l:next_mode = 2
+        elseif index(s:c_count, l:command) != -1
+            let l:next_mode = a:commandset_index
+        elseif index(s:c_motion, l:command) != -1
+            if l:command == 't' || l:command == 'f' || l:command == "'"
+                " prompt single key
+                let l:next_mode = 2
+            else
+                let l:do_execute = v:true
+                let l:next_mode = 0
+            endif
+        elseif index(s:c_command, l:command) != -1
+            if index(["y","g","d","c"], s:command_backlog[-1:-1]) != -1 && s:command_backlog[-1:-1] != s:command_backlog[-2:-2] && mode() !=? 'v'
+                " need motion or repeated command
+                " Potential for bad state here if disparaging command keys are
+                " entered (i.e. yd), but vim can handle checks for this at exe
+                " And checking for cases like y123d would complicate things
+                let l:next_mode = 1
+            elseif index(["i","a","c", "o", "s"], l:command) != -1 || s:command_backlog[-1:-1] ==# 'R'
+                "'Insert' mode, do general transcription
+                let l:req = {"method": "unguided", "params": a:params}
+                let l:req.params.timestamp = a:msg.result.timestamp
+                let l:req.params.no_context = v:true
+                let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [function("s:subTranProg"), function("s:subTranFinish", [a:params])])})
+                return
+            elseif l:command == 'r' || l:command == 'm'
+                let l:next_mode = 2
+            elseif l:command == '.'
+                let l:next_mode = 0
+                let l:do_execute = v:true
+                let s:command_backlog = s:command_backlog[0:-2] .. s:repeat_command
+            else
+                if l:command ==? 'v'
+                    let l:next_mode = 1
+                else
+                    let l:next_mode = 0
+                endif
+                let l:do_execute = v:true
+            endif
+        else
+            throw "Invalid command state: " .. l:command .. " " .. a:commandset_index .. " " .. s:command_backlog
+        endif
+    endif
+    if l:do_execute
+        if mode() ==?'v' && l:next_mode == 0
+            let l:next_mode = 1
+        elseif match(s:command_backlog, 'c') != -1
+            let l:req = {"method": "unguided", "params": a:params}
+            let l:req.params.timestamp = a:msg.result.timestamp
+            let l:req.params.no_context = v:true
+            let resp = ch_sendexpr(g:lsp_job, req, {"callback": function("s:transcriptionCallback", [function("s:subTranProg"), function("s:subTranFinish", [a:params])])})
+            return
+        endif
+        exe "normal" s:command_backlog
+        if index(s:c_motion + ["u"],l:command) == -1
+            exe "normal a\<C-G>u"
+            let s:repeat_command = s:command_backlog
+            call s:logCallback(0, s:command_backlog)
+        endif
+        let s:command_backlog = ""
+    endif
+    let l:req = {"method": "guided", "params": a:params}
+    let l:req.params.timestamp = a:msg.result.timestamp
+    let l:req.params.commandset_index = l:next_mode
+    let resp = ch_sendexpr(g:lsp_job, l:req, {"callback": function("s:commandCallback",[a:params, l:next_mode])})
+endfunction
+
+func s:loadedCallback(channel, msg)
+    echo "Loading complete"
+    call s:logCallback(a:channel, a:msg)
+endfunction
+
+func s:registerCommandset(commandlist, is_final)
+    let req = {"method": "registerCommandset"}
+    let req.params = a:commandlist
+    call s:logCallback(0, join(a:commandlist))
+    call add(g:whisper_commandlist_spoken, a:commandlist)
+    if a:is_final
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": "s:loadedCallback"})
+    else
+        let resp = ch_sendexpr(g:lsp_job, req, {"callback": "s:logCallback"})
+    endif
+endfunction
+
+func s:registerAllCommands()
+    let l:normal = s:c_special_always + s:c_special_normal + s:c_count + s:c_command + s:c_motion + keys(s:c_user)
+    let l:visual = s:c_special_always + s:c_count + s:c_command + s:c_motion
+    " Currently the same as visual.
+    " let l:post_command = s:c_special_always + s:c_count + s:c_command + s:c_motion
+    let l:single_key = s:c_special_always + split(s:c_lowerkeys, '\zs')
+    let l:area = s:c_special_always + s:c_area
+
+    " Used only for compatibility with the testing script
+    let g:whisper_commandlist_spoken = []
+
+    let s:commandset_list = [l:normal, l:visual, l:single_key, l:area]
+    call s:registerCommandset(s:commandsetToSpoken(l:normal, 0), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:visual, 1), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:single_key, 2), v:false)
+    call s:registerCommandset(s:commandsetToSpoken(l:area, 3), v:true)
+endfunction
+
+func s:commandsetToSpoken(commandset, spoken_index)
+    let l:spoken_list = []
+    for l:command in a:commandset
+        if has_key(s:spoken_dict, l:command)
+            let l:spoken_value = s:spoken_dict[l:command]
+            if type(l:spoken_value) == v:t_dict
+                if has_key(l:spoken_value, a:spoken_index)
+                    let l:spoken_value = l:spoken_value[a:spoken_index]
+                else
+                    if a:spoken_index == 2
+                        let l:spoken_value = l:command
+                    else
+                        let l:spoken_value = l:spoken_value[0]
+                    endif
+                endif
+            else
+                if a:spoken_index == 2
+                    let l:spoken_value = l:command
+                endif
+            endif
+        else
+            let l:spoken_value = l:command
+        endif
+        call add(l:spoken_list, l:spoken_value)
+    endfor
+    return l:spoken_list
+endfunction
+
+" TODO: Check lifetime. If the script is resourced, is the existing
+" s:lsp_job dropped and therefore killed?
+" This seems to not be the case and I've had to deal with zombie processes
+" that survive exiting vim, even though said behavior conflicts with my
+" understanding of the provided documentation
+let s:lsp_opts = {"in_mode": "lsp", "out_mode": "lsp", "err_mode": "nl", "err_io": "buffer", "err_buf": s:output_buffer}
+if !exists("g:lsp_job")
+    if exists("g:whisper_user_commands")
+        let s:c_user = g:whisper_user_commands
+    else
+        let s:c_user = {}
+    endif
+    let g:lsp_job = job_start(s:lsp_command, s:lsp_opts)
+    if job_status(g:lsp_job) == "fail"
+        echoerr "Failed to start whisper job"
+    endif
+    call s:registerAllCommands()
+endif

examples/main/main.cpp

@@ -10,6 +10,10 @@
 #include <vector>
 #include <cstring>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
 const std::vector<std::string> k_colors = {
@@ -55,37 +59,48 @@ struct whisper_params {
     int32_t offset_t_ms  =  0;
     int32_t offset_n     =  0;
     int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
     int32_t max_context  = -1;
     int32_t max_len      =  0;
-    int32_t best_of      =  2;
-    int32_t beam_size    = -1;
+    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
 
     float word_thold    =  0.01f;
     float entropy_thold =  2.40f;
     float logprob_thold = -1.00f;
 
-    bool speed_up       = false;
-    bool translate      = false;
-    bool detect_language= false;
-    bool diarize        = false;
-    bool split_on_word  = false;
-    bool no_fallback    = false;
-    bool output_txt     = false;
-    bool output_vtt     = false;
-    bool output_srt     = false;
-    bool output_wts     = false;
-    bool output_csv     = false;
-    bool output_jsn     = false;
-    bool output_lrc     = false;
-    bool print_special  = false;
-    bool print_colors   = false;
-    bool print_progress = false;
-    bool no_timestamps  = false;
+    bool speed_up        = false;
+    bool debug_mode      = false;
+    bool translate       = false;
+    bool detect_language = false;
+    bool diarize         = false;
+    bool tinydiarize     = false;
+    bool split_on_word   = false;
+    bool no_fallback     = false;
+    bool output_txt      = false;
+    bool output_vtt      = false;
+    bool output_srt      = false;
+    bool output_wts      = false;
+    bool output_csv      = false;
+    bool output_jsn      = false;
+    bool output_jsn_full = false;
+    bool output_lrc      = false;
+    bool print_special   = false;
+    bool print_colors    = false;
+    bool print_progress  = false;
+    bool no_timestamps   = false;
+    bool log_score       = false;
+    bool use_gpu         = true;
 
-    std::string language = "en";
+    std::string language  = "en";
     std::string prompt;
     std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
-    std::string model    = "models/ggml-base.en.bin";
+    std::string model     = "models/ggml-base.en.bin";
+
+    // [TDRZ] speaker turn string
+    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
+
+    std::string openvino_encode_device = "CPU";
 
     std::vector<std::string> fname_inp = {};
     std::vector<std::string> fname_out = {};
@@ -111,41 +126,47 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"    || arg == "--threads")        { params.n_threads      = std::stoi(argv[++i]); }
-        else if (arg == "-p"    || arg == "--processors")     { params.n_processors   = std::stoi(argv[++i]); }
-        else if (arg == "-ot"   || arg == "--offset-t")       { params.offset_t_ms    = std::stoi(argv[++i]); }
-        else if (arg == "-on"   || arg == "--offset-n")       { params.offset_n       = std::stoi(argv[++i]); }
-        else if (arg == "-d"    || arg == "--duration")       { params.duration_ms    = std::stoi(argv[++i]); }
-        else if (arg == "-mc"   || arg == "--max-context")    { params.max_context    = std::stoi(argv[++i]); }
-        else if (arg == "-ml"   || arg == "--max-len")        { params.max_len        = std::stoi(argv[++i]); }
-        else if (arg == "-bo"   || arg == "--best-of")        { params.best_of        = std::stoi(argv[++i]); }
-        else if (arg == "-bs"   || arg == "--beam-size")      { params.beam_size      = std::stoi(argv[++i]); }
-        else if (arg == "-wt"   || arg == "--word-thold")     { params.word_thold     = std::stof(argv[++i]); }
-        else if (arg == "-et"   || arg == "--entropy-thold")  { params.entropy_thold  = std::stof(argv[++i]); }
-        else if (arg == "-lpt"  || arg == "--logprob-thold")  { params.logprob_thold  = std::stof(argv[++i]); }
-        else if (arg == "-su"   || arg == "--speed-up")       { params.speed_up       = true; }
-        else if (arg == "-tr"   || arg == "--translate")      { params.translate      = true; }
-        else if (arg == "-di"   || arg == "--diarize")        { params.diarize        = true; }
-        else if (arg == "-sow"  || arg == "--split-on-word")  { params.split_on_word  = true; }
-        else if (arg == "-nf"   || arg == "--no-fallback")    { params.no_fallback    = true; }
-        else if (arg == "-otxt" || arg == "--output-txt")     { params.output_txt     = true; }
-        else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
-        else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
-        else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
-        else if (arg == "-olrc" || arg == "--output-lrc")     { params.output_lrc     = true; }
-        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
-        else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
-        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
-        else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
-        else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
-        else if (arg == "-pc"   || arg == "--print-colors")   { params.print_colors   = true; }
-        else if (arg == "-pp"   || arg == "--print-progress") { params.print_progress = true; }
-        else if (arg == "-nt"   || arg == "--no-timestamps")  { params.no_timestamps  = true; }
-        else if (arg == "-l"    || arg == "--language")       { params.language       = argv[++i]; }
-        else if (arg == "-dl"   || arg == "--detect-language"){ params.detect_language= true; }
-        else if (                  arg == "--prompt")         { params.prompt         = argv[++i]; }
-        else if (arg == "-m"    || arg == "--model")          { params.model          = argv[++i]; }
-        else if (arg == "-f"    || arg == "--file")           { params.fname_inp.emplace_back(argv[++i]); }
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
+        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
+        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
+        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
+        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
+        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
+        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
+        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
+        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
+        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
+        else if (arg == "-otxt" || arg == "--output-txt")      { params.output_txt      = true; }
+        else if (arg == "-ovtt" || arg == "--output-vtt")      { params.output_vtt      = true; }
+        else if (arg == "-osrt" || arg == "--output-srt")      { params.output_srt      = true; }
+        else if (arg == "-owts" || arg == "--output-words")    { params.output_wts      = true; }
+        else if (arg == "-olrc" || arg == "--output-lrc")      { params.output_lrc      = true; }
+        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
+        else if (arg == "-ocsv" || arg == "--output-csv")      { params.output_csv      = true; }
+        else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
+        else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
+        else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
+        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
+        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
+        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
+        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
+        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
+        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
+        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
+        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -175,9 +196,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
-    fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
     fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
     fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
     fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
     fprintf(stderr, "  -otxt,     --output-txt        [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
     fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
@@ -187,6 +210,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
     fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
     fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
+    fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
@@ -197,6 +221,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
     fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
+    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
+    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
     fprintf(stderr, "\n");
 }
 
@@ -204,8 +231,50 @@ struct whisper_print_user_data {
     const whisper_params * params;
 
     const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
 };
 
+std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+    std::string speaker = "";
+    const int64_t n_samples = pcmf32s[0].size();
+
+    const int64_t is0 = timestamp_to_sample(t0, n_samples);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples);
+
+    double energy0 = 0.0f;
+    double energy1 = 0.0f;
+
+    for (int64_t j = is0; j < is1; j++) {
+        energy0 += fabs(pcmf32s[0][j]);
+        energy1 += fabs(pcmf32s[1][j]);
+    }
+
+    if (energy0 > 1.1*energy1) {
+        speaker = "0";
+    } else if (energy1 > 1.1*energy0) {
+        speaker = "1";
+    } else {
+        speaker = "?";
+    }
+
+    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
+
+    if (!id_only) {
+        speaker.insert(0, "(speaker ");
+        speaker.append(")");
+    }
+
+    return speaker;
+}
+void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
+    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
+    if (progress >= *progress_prev + progress_step) {
+        *progress_prev += progress_step;
+        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
+    }
+}
+
 void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
@@ -235,28 +304,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
         }
 
         if (params.diarize && pcmf32s.size() == 2) {
-            const int64_t n_samples = pcmf32s[0].size();
-
-            const int64_t is0 = timestamp_to_sample(t0, n_samples);
-            const int64_t is1 = timestamp_to_sample(t1, n_samples);
-
-            double energy0 = 0.0f;
-            double energy1 = 0.0f;
-
-            for (int64_t j = is0; j < is1; j++) {
-                energy0 += fabs(pcmf32s[0][j]);
-                energy1 += fabs(pcmf32s[1][j]);
-            }
-
-            if (energy0 > 1.1*energy1) {
-                speaker = "(speaker 0)";
-            } else if (energy1 > 1.1*energy0) {
-                speaker = "(speaker 1)";
-            } else {
-                speaker = "(speaker ?)";
-            }
-
-            //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, %s\n", is0, is1, energy0, energy1, speaker.c_str());
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
         }
 
         if (params.print_colors) {
@@ -281,6 +329,12 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
             printf("%s%s", speaker.c_str(), text);
         }
 
+        if (params.tinydiarize) {
+            if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                printf("%s", params.tdrz_speaker_turn.c_str());
+            }
+        }
+
         // with timestamps or speakers: each segment on new line
         if (!params.no_timestamps || params.diarize) {
             printf("\n");
@@ -290,7 +344,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
     }
 }
 
-bool output_txt(struct whisper_context * ctx, const char * fname) {
+bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -302,13 +356,22 @@ bool output_txt(struct whisper_context * ctx, const char * fname) {
     const int n_segments = whisper_full_n_segments(ctx);
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
-        fout << text << "\n";
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        fout << speaker << text << "\n";
     }
 
     return true;
 }
 
-bool output_vtt(struct whisper_context * ctx, const char * fname) {
+bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -324,15 +387,23 @@ bool output_vtt(struct whisper_context * ctx, const char * fname) {
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1, true);
+            speaker.insert(0, "<v Speaker");
+            speaker.append(">");
+        }
 
         fout << to_timestamp(t0) << " --> " << to_timestamp(t1) << "\n";
-        fout << text << "\n\n";
+        fout << speaker << text << "\n\n";
     }
 
     return true;
 }
 
-bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -346,10 +417,16 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
 
         fout << i + 1 + params.offset_n << "\n";
         fout << to_timestamp(t0, true) << " --> " << to_timestamp(t1, true) << "\n";
-        fout << text << "\n\n";
+        fout << speaker << text << "\n\n";
     }
 
     return true;
@@ -386,7 +463,7 @@ char *escape_double_quotes_and_backslashes(const char *str) {
     return escaped;
 }
 
-bool output_csv(struct whisper_context * ctx, const char * fname) {
+bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -396,7 +473,13 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
     const int n_segments = whisper_full_n_segments(ctx);
-    fout << "start,end,text\n";
+    fout << "start,end,";
+    if (params.diarize && pcmf32s.size() == 2)
+    {
+        fout << "speaker,";
+    }
+    fout << "text\n";
+
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
@@ -404,13 +487,42 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
         char * text_escaped = escape_double_quotes_and_backslashes(text);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
-        fout << 10 * t0 << "," << 10 * t1 << ",\"" << text_escaped    << "\"\n";
+        fout << 10 * t0 << "," << 10 * t1 << ",";
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            fout << estimate_diarization_speaker(pcmf32s, t0, t1, true) << ",";
+        }
+        fout << "\"" << text_escaped << "\"\n";
     }
 
     return true;
 }
 
-bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+    std::ofstream fout(fname);
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    // fprintf(stderr,"segments: %d\n",n_segments);
+    for (int i = 0; i < n_segments; ++i) {
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        // fprintf(stderr,"tokens: %d\n",n_tokens);
+        for (int j = 0; j < n_tokens; j++) {
+            auto token = whisper_full_get_token_text(ctx, i, j);
+            auto probability = whisper_full_get_token_p(ctx, i, j);
+            fout << token << '\t' << probability << std::endl;
+            // fprintf(stderr,"token: %s %f\n",token,probability);
+	    }
+    }
+    return true;
+}
+
+bool output_json(
+             struct whisper_context * ctx,
+                         const char * fname,
+               const whisper_params & params,
+    std::vector<std::vector<float>>   pcmf32s,
+                               bool   full) {
     std::ofstream fout(fname);
     int indent = 0;
 
@@ -424,13 +536,13 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
         indent++;
     };
 
-    auto end_arr = [&](bool end = false) {
+    auto end_arr = [&](bool end) {
         indent--;
         doindent();
-        fout << (end ? "]\n" : "},\n");
+        fout << (end ? "]\n" : "],\n");
     };
 
-    auto start_obj = [&](const char *name = nullptr) {
+    auto start_obj = [&](const char *name) {
         doindent();
         if (name) {
             fout << "\"" << name << "\": {\n";
@@ -440,7 +552,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
         indent++;
     };
 
-    auto end_obj = [&](bool end = false) {
+    auto end_obj = [&](bool end) {
         indent--;
         doindent();
         fout << (end ? "}\n" : "},\n");
@@ -451,82 +563,119 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
         fout << "\"" << name << "\": ";
     };
 
-    auto value_s = [&](const char *name, const char *val, bool end = false) {
+    auto value_s = [&](const char *name, const char *val, bool end) {
         start_value(name);
         char * val_escaped = escape_double_quotes_and_backslashes(val);
         fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
         free(val_escaped);
     };
 
-    auto end_value = [&](bool end = false) {
+    auto end_value = [&](bool end) {
         fout << (end ? "\n" : ",\n");
     };
 
-    auto value_i = [&](const char *name, const int64_t val, bool end = false) {
+    auto value_i = [&](const char *name, const int64_t val, bool end) {
         start_value(name);
         fout << val;
         end_value(end);
     };
 
-    auto value_b = [&](const char *name, const bool val, bool end = false) {
+    auto value_f = [&](const char *name, const float val, bool end) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end) {
         start_value(name);
         fout << (val ? "true" : "false");
         end_value(end);
     };
 
+    auto times_o = [&](int64_t t0, int64_t t1, bool end) {
+        start_obj("timestamps");
+        value_s("from", to_timestamp(t0, true).c_str(), false);
+        value_s("to", to_timestamp(t1, true).c_str(), true);
+        end_obj(false);
+        start_obj("offsets");
+        value_i("from", t0 * 10, false);
+        value_i("to", t1 * 10, true);
+        end_obj(end);
+    };
+
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
         return false;
     }
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
-    start_obj();
-        value_s("systeminfo", whisper_print_system_info());
+    start_obj(nullptr);
+        value_s("systeminfo", whisper_print_system_info(), false);
         start_obj("model");
-            value_s("type", whisper_model_type_readable(ctx));
-            value_b("multilingual", whisper_is_multilingual(ctx));
-            value_i("vocab", whisper_model_n_vocab(ctx));
+            value_s("type", whisper_model_type_readable(ctx), false);
+            value_b("multilingual", whisper_is_multilingual(ctx), false);
+            value_i("vocab", whisper_model_n_vocab(ctx), false);
             start_obj("audio");
-                value_i("ctx", whisper_model_n_audio_ctx(ctx));
-                value_i("state", whisper_model_n_audio_state(ctx));
-                value_i("head", whisper_model_n_audio_head(ctx));
+                value_i("ctx", whisper_model_n_audio_ctx(ctx), false);
+                value_i("state", whisper_model_n_audio_state(ctx), false);
+                value_i("head", whisper_model_n_audio_head(ctx), false);
                 value_i("layer", whisper_model_n_audio_layer(ctx), true);
-            end_obj();
+            end_obj(false);
             start_obj("text");
-                value_i("ctx", whisper_model_n_text_ctx(ctx));
-                value_i("state", whisper_model_n_text_state(ctx));
-                value_i("head", whisper_model_n_text_head(ctx));
+                value_i("ctx", whisper_model_n_text_ctx(ctx), false);
+                value_i("state", whisper_model_n_text_state(ctx), false);
+                value_i("head", whisper_model_n_text_head(ctx), false);
                 value_i("layer", whisper_model_n_text_layer(ctx), true);
-            end_obj();
-            value_i("mels", whisper_model_n_mels(ctx));
+            end_obj(false);
+            value_i("mels", whisper_model_n_mels(ctx), false);
             value_i("ftype", whisper_model_ftype(ctx), true);
-        end_obj();
+        end_obj(false);
         start_obj("params");
-            value_s("model", params.model.c_str());
-            value_s("language", params.language.c_str());
+            value_s("model", params.model.c_str(), false);
+            value_s("language", params.language.c_str(), false);
             value_b("translate", params.translate, true);
-        end_obj();
+        end_obj(false);
         start_obj("result");
             value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
-        end_obj();
+        end_obj(false);
         start_arr("transcription");
 
             const int n_segments = whisper_full_n_segments(ctx);
             for (int i = 0; i < n_segments; ++i) {
                 const char * text = whisper_full_get_segment_text(ctx, i);
+
                 const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                 const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-                start_obj();
-                    start_obj("timestamps");
-                        value_s("from", to_timestamp(t0, true).c_str());
-                        value_s("to", to_timestamp(t1, true).c_str(), true);
-                    end_obj();
-                    start_obj("offsets");
-                        value_i("from", t0 * 10);
-                        value_i("to", t1 * 10, true);
-                    end_obj();
-                    value_s("text", text, true);
+                start_obj(nullptr);
+                    times_o(t0, t1, false);
+                    value_s("text", text, !params.diarize && !params.tinydiarize && !full);
+
+                    if (full) {
+                        start_arr("tokens");
+                        const int n = whisper_full_n_tokens(ctx, i);
+                        for (int j = 0; j < n; ++j) {
+                            auto token = whisper_full_get_token_data(ctx, i, j);
+                            start_obj(nullptr);
+                                value_s("text", whisper_token_to_str(ctx, token.id), false);
+                                if(token.t0 > -1 && token.t1 > -1) {
+                                    // If we have per-token timestamps, write them out
+                                    times_o(token.t0, token.t1, false);
+                                }
+                                value_i("id", token.id, false);
+                                value_f("p", token.p, true);
+                            end_obj(j == (n - 1));
+                        }
+                        end_arr(!params.diarize && !params.tinydiarize);
+                    }
+
+                    if (params.diarize && pcmf32s.size() == 2) {
+                        value_s("speaker", estimate_diarization_speaker(pcmf32s, t0, t1, true).c_str(), true);
+                    }
+
+                    if (params.tinydiarize) {
+                        value_b("speaker_turn_next", whisper_full_get_segment_speaker_turn_next(ctx, i), true);
+                    }
                 end_obj(i == (n_segments - 1));
             }
 
@@ -538,7 +687,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@@ -575,6 +724,11 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
         fout << "drawtext=fontfile='" << font << "':fontsize=24:fontcolor=gray:x=(w-text_w)/2:y=h/2:text='':enable='between(t," << t0/100.0 << "," << t0/100.0 << ")'";
 
         bool is_first = true;
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
 
         for (int j = 0; j < n; ++j) {
             const auto & token = tokens[j];
@@ -583,13 +737,19 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
                 continue;
             }
 
-            std::string txt_bg;
-            std::string txt_fg; // highlight token
-            std::string txt_ul; // underline
+            std::string txt_bg = "";
+            std::string txt_fg = ""; // highlight token
+            std::string txt_ul = ""; // underline
 
-            txt_bg = "> ";
-            txt_fg = "> ";
-            txt_ul = "\\ \\ ";
+            if (params.diarize && pcmf32s.size() == 2) {
+                txt_bg = speaker;
+                txt_fg = speaker;
+                txt_ul = "\\ \\ \\ \\ \\ \\ \\ \\ \\ \\ \\ ";
+            }
+
+            txt_bg.append("> ");
+            txt_fg.append("> ");
+            txt_ul.append("\\ \\ ");
 
             {
                 for (int k = 0; k < n; ++k) {
@@ -652,8 +812,7 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
     return true;
 }
 
-bool output_lrc(struct whisper_context * ctx, const char * fname) {
-
+bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@@ -678,8 +837,16 @@ bool output_lrc(struct whisper_context * ctx, const char * fname) {
         char buf[16];
         snprintf(buf, sizeof(buf), "%02d:%02d.%02d", (int) min, (int) sec, (int) ( msec / 10));
         std::string timestamp_lrc = std::string(buf);
+        std::string speaker = "";
 
-        fout <<  '[' << timestamp_lrc << ']' << text << "\n";
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        fout <<  '[' << timestamp_lrc << ']' << speaker << text << "\n";
     }
 
     return true;
@@ -689,6 +856,7 @@ int main(int argc, char ** argv) {
     whisper_params params;
 
     if (whisper_params_parse(argc, argv, params) == false) {
+        whisper_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -704,15 +872,27 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
+    if (params.diarize && params.tinydiarize) {
+        fprintf(stderr, "error: cannot use both --diarize and --tinydiarize\n");
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
     // whisper init
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     if (ctx == nullptr) {
         fprintf(stderr, "error: failed to initialize whisper context\n");
         return 3;
     }
 
+    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -745,11 +925,12 @@ int main(int argc, char ** argv) {
             if (params.detect_language) {
                 params.language = "auto";
             }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
-                    params.n_threads, params.n_processors,
+                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
+                    params.tinydiarize ? "tdrz = 1, " : "",
                     params.no_timestamps ? 0 : 1);
 
             fprintf(stderr, "\n");
@@ -773,12 +954,15 @@ int main(int argc, char ** argv) {
             wparams.offset_ms        = params.offset_t_ms;
             wparams.duration_ms      = params.duration_ms;
 
-            wparams.token_timestamps = params.output_wts || params.max_len > 0;
+            wparams.token_timestamps = params.output_wts || params.output_jsn_full || params.max_len > 0;
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
             wparams.split_on_word    = params.split_on_word;
 
             wparams.speed_up         = params.speed_up;
+            wparams.debug_mode       = params.debug_mode;
+
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
 
             wparams.initial_prompt   = params.prompt.c_str();
 
@@ -789,7 +973,7 @@ int main(int argc, char ** argv) {
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
 
-            whisper_print_user_data user_data = { &params, &pcmf32s };
+            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
@@ -797,8 +981,14 @@ int main(int argc, char ** argv) {
                 wparams.new_segment_callback_user_data = &user_data;
             }
 
-            // example for abort mechanism
-            // in this example, we do not abort the processing, but we could if the flag is set to true
+            if (wparams.print_progress) {
+                wparams.progress_callback           = whisper_print_progress_callback;
+                wparams.progress_callback_user_data = &user_data;
+            }
+
+            // examples for abort mechanism
+            // in examples below, we do not abort the processing, but we could if the flag is set to true
+
             // the callback is called before every encoder run - if it returns false, the processing is aborted
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
@@ -810,6 +1000,17 @@ int main(int argc, char ** argv) {
                 wparams.encoder_begin_callback_user_data = &is_aborted;
             }
 
+            // the callback is called before every computation - if it returns true, the computation is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.abort_callback = [](void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return is_aborted;
+                };
+                wparams.abort_callback_user_data = &is_aborted;
+            }
+
             if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                 return 10;
@@ -823,43 +1024,49 @@ int main(int argc, char ** argv) {
             // output to text file
             if (params.output_txt) {
                 const auto fname_txt = fname_out + ".txt";
-                output_txt(ctx, fname_txt.c_str());
+                output_txt(ctx, fname_txt.c_str(), params, pcmf32s);
             }
 
             // output to VTT file
             if (params.output_vtt) {
                 const auto fname_vtt = fname_out + ".vtt";
-                output_vtt(ctx, fname_vtt.c_str());
+                output_vtt(ctx, fname_vtt.c_str(), params, pcmf32s);
             }
 
             // output to SRT file
             if (params.output_srt) {
                 const auto fname_srt = fname_out + ".srt";
-                output_srt(ctx, fname_srt.c_str(), params);
+                output_srt(ctx, fname_srt.c_str(), params, pcmf32s);
             }
 
             // output to WTS file
             if (params.output_wts) {
                 const auto fname_wts = fname_out + ".wts";
-                output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);
+                output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE, pcmf32s);
             }
 
             // output to CSV file
             if (params.output_csv) {
                 const auto fname_csv = fname_out + ".csv";
-                output_csv(ctx, fname_csv.c_str());
+                output_csv(ctx, fname_csv.c_str(), params, pcmf32s);
             }
 
             // output to JSON file
             if (params.output_jsn) {
                 const auto fname_jsn = fname_out + ".json";
-                output_json(ctx, fname_jsn.c_str(), params);
+                output_json(ctx, fname_jsn.c_str(), params, pcmf32s, params.output_jsn_full);
             }
 
             // output to LRC file
             if (params.output_lrc) {
                 const auto fname_lrc = fname_out + ".lrc";
-                output_lrc(ctx, fname_lrc.c_str());
+                output_lrc(ctx, fname_lrc.c_str(), params, pcmf32s);
+            }
+
+            // output to score file
+            if (params.log_score) {
+                const auto fname_score = fname_out + ".score.txt";
+                output_score(ctx, fname_score.c_str(), params, pcmf32s);
             }
         }
     }

examples/quantize/quantize.cpp

@@ -57,7 +57,7 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f
     {
         uint32_t magic;
         finp.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6c) {
+        if (magic != GGML_FILE_MAGIC) {
             fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
             return false;
         }
@@ -99,17 +99,17 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f
         fprintf(stderr, "%s: ftype (dst)   = %d\n", __func__, ftype_dst);
         fprintf(stderr, "%s: qntvr (dst)   = %d\n", __func__, GGML_QNT_VERSION);
 
-        fout.write((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
-        fout.write((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
-        fout.write((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
-        fout.write((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
-        fout.write((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
-        fout.write((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
-        fout.write((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
-        fout.write((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
-        fout.write((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
-        fout.write((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
-        fout.write((char *) &ftype_dst,             sizeof(hparams.ftype));
+        fout.write((const char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        fout.write((const char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        fout.write((const char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        fout.write((const char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        fout.write((const char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        fout.write((const char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        fout.write((const char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        fout.write((const char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        fout.write((const char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        fout.write((const char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        fout.write((const char *) &ftype_dst,             sizeof(hparams.ftype));
     }
 
     // load mel filters
@@ -138,15 +138,17 @@ bool whisper_model_quantize(const std::string & fname_inp, const std::string & f
         //    return false;
         //}
 
-        std::string word;
+        char word[129];
+
         for (int i = 0; i < n_vocab; i++) {
             uint32_t len;
             finp.read ((char *) &len, sizeof(len));
             fout.write((char *) &len, sizeof(len));
 
-            word.resize(len);
-            finp.read ((char *) word.data(), len);
-            fout.write((char *) word.data(), len);
+            word[len] = '\0';
+
+            finp.read ((char *) word, len);
+            fout.write((char *) word, len);
 
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;

examples/server/CMakeLists.txt

@@ -0,0 +1,6 @@
+set(TARGET server)
+add_executable(${TARGET} server.cpp httplib.h json.hpp)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/server/README.md

@@ -0,0 +1,59 @@
+# whisper.cpp http server
+
+Simple http server. WAV Files are passed to the inference model via http requests.
+
+```
+./server -h
+
+usage: ./bin/server [options]
+
+options:
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -sow,      --split-on-word     [false  ] split on word rather than on token
+  -bo N,     --best-of N         [2      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -tr,       --translate         [false  ] translate from source language to english
+  -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
+  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  --host HOST,                   [127.0.0.1] Hostname/ip-adress for the server
+  --port PORT,                   [8080   ] Port number for the server
+```
+
+## request examples
+
+**/inference**
+```
+curl 127.0.0.1:8080/inference \
+-H "Content-Type: multipart/form-data" \
+-F file="@<file-path>" \
+-F temperature="0.2" \
+-F response-format="json"
+```
+
+**/load**
+```
+curl 127.0.0.1:8080/load \
+-H "Content-Type: multipart/form-data" \
+-F model="<path-to-model-file>"
+```

examples/server/server.cpp

@@ -0,0 +1,699 @@
+#include "common.h"
+
+#include "whisper.h"
+#include "httplib.h"
+#include "json.hpp"
+
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+#include <cstring>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+using namespace httplib;
+using json = nlohmann::json;
+
+namespace {
+
+// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
+// Lowest is red, middle is yellow, highest is green.
+const std::vector<std::string> k_colors = {
+    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
+    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
+};
+
+// output formats
+const std::string json_format   = "json";
+const std::string text_format   = "text";
+const std::string srt_format    = "srt";
+const std::string vjson_format  = "verbose_json";
+const std::string vtt_format    = "vtt";
+
+struct server_params
+{
+    std::string hostname = "127.0.0.1";
+    std::string public_path = "examples/server/public";
+
+    int32_t port          = 8080;
+    int32_t read_timeout  = 600;
+    int32_t write_timeout = 600;
+};
+
+struct whisper_params {
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
+    int32_t max_context  = -1;
+    int32_t max_len      =  0;
+    int32_t best_of      =  2;
+    int32_t beam_size    = -1;
+
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;
+    float userdef_temp  =  0.20f;
+
+    bool speed_up        = false;
+    bool debug_mode      = false;
+    bool translate       = false;
+    bool detect_language = false;
+    bool diarize         = false;
+    bool tinydiarize     = false;
+    bool split_on_word   = false;
+    bool no_fallback     = false;
+    bool print_special   = false;
+    bool print_colors    = false;
+    bool print_progress  = false;
+    bool no_timestamps   = false;
+    bool use_gpu         = true;
+
+    std::string language        = "en";
+    std::string prompt          = "";
+    std::string font_path       = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    std::string model           = "models/ggml-base.en.bin";
+
+    std::string response_format     = json_format;
+
+    // [TDRZ] speaker turn string
+    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
+
+    std::string openvino_encode_device = "CPU";
+};
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
+bool is_file_exist(const char *fileName)
+{
+    std::ifstream infile(fileName);
+    return infile.good();
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params,
+                         const server_params& sparams) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options] \n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    // fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
+    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
+    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
+    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
+    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    // server params
+    fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
+    fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
+    fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
+    fprintf(stderr, "\n");
+}
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params, server_params & sparams) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params, sparams);
+            exit(0);
+        }
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
+        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
+        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
+        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
+        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
+        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
+        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
+        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
+        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
+        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
+        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
+        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
+        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
+        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
+        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
+        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
+        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
+        // server params
+        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
+        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
+        else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params, sparams);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
+};
+
+std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+    std::string speaker = "";
+    const int64_t n_samples = pcmf32s[0].size();
+
+    const int64_t is0 = timestamp_to_sample(t0, n_samples);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples);
+
+    double energy0 = 0.0f;
+    double energy1 = 0.0f;
+
+    for (int64_t j = is0; j < is1; j++) {
+        energy0 += fabs(pcmf32s[0][j]);
+        energy1 += fabs(pcmf32s[1][j]);
+    }
+
+    if (energy0 > 1.1*energy1) {
+        speaker = "0";
+    } else if (energy1 > 1.1*energy0) {
+        speaker = "1";
+    } else {
+        speaker = "?";
+    }
+
+    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
+
+    if (!id_only) {
+        speaker.insert(0, "(speaker ");
+        speaker.append(")");
+    }
+
+    return speaker;
+}
+
+void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
+    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
+    if (progress >= *progress_prev + progress_step) {
+        *progress_prev += progress_step;
+        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
+    }
+}
+
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+
+    std::string speaker = "";
+
+    int64_t t0 = 0;
+    int64_t t1 = 0;
+
+    // print the last n_new segments
+    const int s0 = n_segments - n_new;
+
+    if (s0 == 0) {
+        printf("\n");
+    }
+
+    for (int i = s0; i < n_segments; i++) {
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
+        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        if (params.print_colors) {
+            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
+                if (params.print_special == false) {
+                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
+                    if (id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+                }
+
+                const char * text = whisper_full_get_token_text(ctx, i, j);
+                const float  p    = whisper_full_get_token_p   (ctx, i, j);
+
+                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
+
+                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
+            }
+        } else {
+            const char * text = whisper_full_get_segment_text(ctx, i);
+
+            printf("%s%s", speaker.c_str(), text);
+        }
+
+        if (params.tinydiarize) {
+            if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                printf("%s", params.tdrz_speaker_turn.c_str());
+            }
+        }
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+        fflush(stdout);
+    }
+}
+
+std::string output_str(struct whisper_context * ctx, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::stringstream result;
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        result << speaker << text << "\n";
+    }
+    return result.str();
+}
+
+void get_req_parameters(const Request & req, whisper_params & params)
+{
+    // user model configu.has_fileion
+    if (req.has_file("offset-t"))
+    {
+        params.offset_t_ms = std::stoi(req.get_file_value("offset-t").content);
+    }
+    if (req.has_file("offset-n"))
+    {
+        params.offset_n = std::stoi(req.get_file_value("offset-n").content);
+    }
+    if (req.has_file("duration"))
+    {
+        params.duration_ms = std::stoi(req.get_file_value("duration").content);
+    }
+    if (req.has_file("max-context"))
+    {
+        params.max_context = std::stoi(req.get_file_value("max-context").content);
+    }
+    if (req.has_file("prompt"))
+    {
+        params.prompt = req.get_file_value("prompt").content;
+    }
+    if (req.has_file("response-format"))
+    {
+        params.response_format = req.get_file_value("response-format").content;
+    }
+    if (req.has_file("temerature"))
+    {
+        params.userdef_temp = std::stof(req.get_file_value("temperature").content);
+    }
+}
+
+}  // namespace
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+    server_params sparams;
+
+    std::mutex whisper_mutex;
+
+    if (whisper_params_parse(argc, argv, params, sparams) == false) {
+        whisper_print_usage(argc, argv, params, sparams);
+        return 1;
+    }
+
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params, sparams);
+        exit(0);
+    }
+
+    if (params.diarize && params.tinydiarize) {
+        fprintf(stderr, "error: cannot use both --diarize and --tinydiarize\n");
+        whisper_print_usage(argc, argv, params, sparams);
+        exit(0);
+    }
+
+    // whisper init
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 3;
+    }
+
+    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
+    Server svr;
+
+    std::string const default_content = "<html>hello</html>";
+
+    // this is only called if no index.html is found in the public --path
+    svr.Get("/", [&default_content](const Request &, Response &res){
+        res.set_content(default_content, "text/html");
+        return false;
+    });
+
+    svr.Post("/inference", [&](const Request &req, Response &res){
+        // aquire whisper model mutex lock
+        whisper_mutex.lock();
+
+        // first check user requested fields of the request
+        if (!req.has_file("file"))
+        {
+            fprintf(stderr, "error: no 'file' field in the request\n");
+            const std::string error_resp = "{\"error\":\"no 'file' field in the request\"}";
+            res.set_content(error_resp, "application/json");
+            whisper_mutex.unlock();
+            return;
+        }
+        auto audio_file = req.get_file_value("file");
+
+        // check non-required fields
+        get_req_parameters(req, params);
+
+        std::string filename{audio_file.filename};
+        printf("Received request: %s\n", filename.c_str());
+
+        // audio arrays
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+        // write file to temporary file
+        std::ofstream temp_file{filename, std::ios::binary};
+        temp_file << audio_file.content;
+
+        // read wav content into pcmf32
+        if (!::read_wav(filename, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", filename.c_str());
+            const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+            res.set_content(error_resp, "application/json");
+            whisper_mutex.unlock();
+            return;
+        }
+        // remove temp file
+        std::remove(filename.c_str());
+
+        printf("Successfully loaded %s\n", filename.c_str());
+
+        // print system information
+        {
+            fprintf(stderr, "\n");
+            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
+        }
+
+        // print some info about the processing
+        {
+            fprintf(stderr, "\n");
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            if (params.detect_language) {
+                params.language = "auto";
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
+                    __func__, filename.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
+                    params.n_threads, params.n_processors,
+                    params.language.c_str(),
+                    params.translate ? "translate" : "transcribe",
+                    params.tinydiarize ? "tdrz = 1, " : "",
+                    params.no_timestamps ? 0 : 1);
+
+            fprintf(stderr, "\n");
+        }
+
+        // run the inference
+        {
+
+            printf("Running whisper.cpp inference on %s\n", filename.c_str());
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+
+            wparams.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.detect_language  = params.detect_language;
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.thold_pt         = params.word_thold;
+            wparams.split_on_word    = params.split_on_word;
+
+            wparams.speed_up         = params.speed_up;
+            wparams.debug_mode       = params.debug_mode;
+
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
+
+            wparams.initial_prompt   = params.prompt.c_str();
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.temperature_inc  = params.userdef_temp;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+
+            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
+
+            // this callback is called on each new segment
+            if (!wparams.print_realtime) {
+                wparams.new_segment_callback           = whisper_print_segment_callback;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            if (wparams.print_progress) {
+                wparams.progress_callback           = whisper_print_progress_callback;
+                wparams.progress_callback_user_data = &user_data;
+            }
+
+            // examples for abort mechanism
+            // in examples below, we do not abort the processing, but we could if the flag is set to true
+
+            // the callback is called before every encoder run - if it returns false, the processing is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
+            }
+
+            // the callback is called before every computation - if it returns true, the computation is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.abort_callback = [](void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return is_aborted;
+                };
+                wparams.abort_callback_user_data = &is_aborted;
+            }
+
+            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                fprintf(stderr, "%s: failed to process audio\n", argv[0]);
+                const std::string error_resp = "{\"error\":\"failed to process audio\"}";
+                res.set_content(error_resp, "application/json");
+                whisper_mutex.unlock();
+                return;
+            }
+        }
+
+        // return results to user
+        if (params.response_format == text_format)
+        {
+            std::string results = output_str(ctx, params, pcmf32s);
+            res.set_content(results.c_str(), "text/html");
+        }
+        // TODO add more output formats
+        else
+        {
+            std::string results = output_str(ctx, params, pcmf32s);
+            json jres = json{
+                {"text", results}
+            };
+            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
+                            "application/json");
+        }
+
+        // return whisper model mutex lock
+        whisper_mutex.unlock();
+    });
+    svr.Post("/load", [&](const Request &req, Response &res){
+        whisper_mutex.lock();
+        if (!req.has_file("model"))
+        {
+            fprintf(stderr, "error: no 'model' field in the request\n");
+            const std::string error_resp = "{\"error\":\"no 'model' field in the request\"}";
+            res.set_content(error_resp, "application/json");
+            whisper_mutex.unlock();
+            return;
+        }
+        std::string model = req.get_file_value("model").content;
+        if (!is_file_exist(model.c_str()))
+        {
+            fprintf(stderr, "error: 'model': %s not found!\n", model.c_str());
+            const std::string error_resp = "{\"error\":\"model not found!\"}";
+            res.set_content(error_resp, "application/json");
+            whisper_mutex.unlock();
+            return;
+        }
+
+        // clean up
+        whisper_free(ctx);
+
+        // whisper init
+        ctx = whisper_init_from_file_with_params(model.c_str(), cparams);
+
+        // TODO perhaps load prior model here instead of exit
+        if (ctx == nullptr) {
+            fprintf(stderr, "error: model init  failed, no model loaded must exit\n");
+            exit(1);
+        }
+
+        // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+        whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
+        const std::string success = "Load was successful!";
+        res.set_content(success, "application/text");
+
+        // check if the model is in the file system
+        whisper_mutex.unlock();
+    });
+
+    svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) {
+        const char fmt[] = "500 Internal Server Error\n%s";
+        char buf[BUFSIZ];
+        try {
+            std::rethrow_exception(std::move(ep));
+        } catch (std::exception &e) {
+            snprintf(buf, sizeof(buf), fmt, e.what());
+        } catch (...) {
+            snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+        }
+        res.set_content(buf, "text/plain");
+        res.status = 500;
+    });
+
+    svr.set_error_handler([](const Request &, Response &res) {
+        if (res.status == 400) {
+            res.set_content("Invalid request", "text/plain");
+        } else if (res.status != 500) {
+            res.set_content("File Not Found", "text/plain");
+            res.status = 404;
+        }
+    });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout(sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
+    {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n",
+                sparams.hostname.c_str(), sparams.port);
+        return 1;
+    }
+
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
+
+    // to make it ctrl+clickable:
+    printf("\nwhisper server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    if (!svr.listen_after_bind())
+    {
+        return 1;
+    }
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}

examples/stream.wasm/emscripten.cpp

@@ -132,7 +132,7 @@ EMSCRIPTEN_BINDINGS(stream) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/stream/README.md

@@ -39,6 +39,20 @@ brew install sdl2
 make stream
 ```
 
+Ensure you are at the root of the repo when running `make stream`.  Not within the `examples/stream` dir
+as the libraries needed like `common-sdl.h` are located within `examples`.  Attempting to compile within
+`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
+
+```bash
+whisper.cpp/examples/stream$ make stream
+g++     stream.cpp   -o stream
+stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
+    6 | #include "common/sdl.h"
+      |          ^~~~~~~~~~~~~~
+compilation terminated.
+make: *** [<builtin>: stream] Error 1
+```
+
 ## Web version
 
 This tool can also run in the browser: [examples/stream.wasm](/examples/stream.wasm)

examples/stream/stream.cpp

@@ -2,9 +2,8 @@
 //
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
 //
-
-#include "common.h"
 #include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
 
 #include <cassert>
@@ -14,6 +13,7 @@
 #include <vector>
 #include <fstream>
 
+
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
 std::string to_timestamp(int64_t t) {
@@ -47,6 +47,9 @@ struct whisper_params {
     bool print_special = false;
     bool no_context    = true;
     bool no_timestamps = false;
+    bool tinydiarize   = false;
+    bool save_audio    = false; // save audio to wav file
+    bool use_gpu       = true;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
@@ -63,23 +66,27 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
-        else if (                 arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
-        else if (                 arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
-        else if (                 arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
-        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
-        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
-        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
-        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
-        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
-        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
-        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-t"    || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (                  arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
+        else if (                  arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
+        else if (                  arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
+        else if (arg == "-c"    || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"   || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
+        else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-kc"   || arg == "--keep-context")  { params.no_context    = false; }
+        else if (arg == "-l"    || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"    || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")   { params.tinydiarize   = true; }
+        else if (arg == "-sa"   || arg == "--save-audio")    { params.save_audio    = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")        { params.use_gpu       = false; }
+
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -113,6 +120,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
     fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
+    fprintf(stderr, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
+    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
     fprintf(stderr, "\n");
 }
 
@@ -150,14 +160,16 @@ int main(int argc, char ** argv) {
     audio.resume();
 
     // whisper init
-
     if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
         fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
         whisper_print_usage(argc, argv, params);
         exit(0);
     }
 
-    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+    struct whisper_context_params cparams;
+    cparams.use_gpu = params.use_gpu;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
 
     std::vector<float> pcmf32    (n_samples_30s, 0.0f);
     std::vector<float> pcmf32_old;
@@ -208,14 +220,28 @@ int main(int argc, char ** argv) {
         }
     }
 
-    printf("[Start speaking]");
+    wav_writer wavWriter;
+    // save wav file
+    if (params.save_audio) {
+        // Get current date/time for filename
+        time_t now = time(0);
+        char buffer[80];
+        strftime(buffer, sizeof(buffer), "%Y%m%d%H%M%S", localtime(&now));
+        std::string filename = std::string(buffer) + ".wav";
+
+        wavWriter.open(filename, WHISPER_SAMPLE_RATE, 16, 1);
+    }
+    printf("[Start speaking]\n");
     fflush(stdout);
 
-          auto t_last  = std::chrono::high_resolution_clock::now();
+    auto t_last  = std::chrono::high_resolution_clock::now();
     const auto t_start = t_last;
 
     // main audio loop
     while (is_running) {
+        if (params.save_audio) {
+            wavWriter.write(pcmf32_new.data(), pcmf32_new.size());
+        }
         // handle Ctrl + C
         is_running = sdl_poll_events();
 
@@ -299,6 +325,8 @@ int main(int argc, char ** argv) {
             wparams.audio_ctx        = params.audio_ctx;
             wparams.speed_up         = params.speed_up;
 
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
+
             // disable temperature fallback
             //wparams.temperature_inc  = -1.0f;
             wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
@@ -344,10 +372,19 @@ int main(int argc, char ** argv) {
                         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
 
-                        printf ("[%s --> %s]  %s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), text);
+                        std::string output = "[" + to_timestamp(t0) + " --> " + to_timestamp(t1) + "]  " + text;
+
+                        if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                            output += " [SPEAKER_TURN]";
+                        }
+
+                        output += "\n";
+
+                        printf("%s", output.c_str());
+                        fflush(stdout);
 
                         if (params.fname_out.length() > 0) {
-                            fout << "[" << to_timestamp(t0) << " --> " << to_timestamp(t1) << "]  " << text << std::endl;
+                            fout << output;
                         }
                     }
                 }
@@ -356,7 +393,7 @@ int main(int argc, char ** argv) {
                     fout << std::endl;
                 }
 
-                if (use_vad){
+                if (use_vad) {
                     printf("\n");
                     printf("### Transcription %d END\n", n_iter);
                 }

examples/talk-llama/CMakeLists.txt

@@ -7,7 +7,16 @@ if (WHISPER_SDL2)
 
     # TODO: this is temporary
     #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+    add_executable(${TARGET}
+        talk-llama.cpp
+        llama.cpp
+        ../common.cpp
+        ../common-sdl.cpp
+        ../../ggml.c
+        ../../ggml-alloc.c
+        ../../ggml-backend.c
+        ../../ggml-quants.c
+        ../../whisper.cpp)
 
     target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
     target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

examples/talk-llama/README.md

@@ -2,6 +2,12 @@
 
 Talk with an LLaMA AI in your terminal
 
+*Latest perf as of 2 Nov 2023 using Whisper Medium + LLaMA v2 13B Q8_0 on M2 Ultra:*
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-60c6b391649e
+
+*Previous demo running on CPUs*
+
 [Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
 
 ## Building
@@ -19,7 +25,7 @@ brew install sdl2
 make talk-llama
 
 # Run it
-./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
 ```
 
 - The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
@@ -36,14 +42,14 @@ This feature is especially helpful for maintaining context in long conversations
 Example usage:
 
 ```bash
-./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
 ```
 
 ## TTS
 
 For best experience, this example needs a TTS tool to convert the generated text responses to voice.
-You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
-By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
+You can use any TTS engine that you would like - simply edit the [speak](speak) script to your needs.
+By default, it is configured to use MacOS's `say` or Windows SpeechSynthesizer, but you can use whatever you wish.
 
 ## Discussion