ios : update SPM package

* whisper : fix bench regression

* ggml : use sched_yield when using BLAS + add comment

.github/workflows/bindings-go.yml

@@ -1,4 +1,4 @@
-name: Bindings Tests
+name: Bindings Tests (Go)
 on:
   push:
     paths:

.github/workflows/bindings-ruby.yml

@@ -0,0 +1,22 @@
+name: Bindings Tests (Ruby)
+on:
+  push:
+    paths:
+      - bindings/ruby/**
+      - whisper.h
+  pull_request:
+    paths:
+      - bindings/ruby/**
+      - whisper.h
+
+jobs:
+  ubuntu-latest:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: ruby/setup-ruby@v1
+        with:
+          ruby-version: '3.0'
+      - uses: actions/checkout@v1
+      - run: |
+          cd bindings/ruby/ext
+          ruby extconf.rb && make

.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 ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ 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 ${{ env.ubuntu_image }} /bin/sh -c '
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ 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
@@ -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
@@ -187,7 +227,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v1
+        uses: actions/checkout@v3
 
       - name: Add msbuild to PATH
         uses: microsoft/setup-msbuild@v1
@@ -236,6 +276,61 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
+  windows-cublas:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [x64]
+        cublas: [ON]
+        sdl2: [ON]
+        include:
+          - arch: x64
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
+
+    steps:
+      - name: Clone
+        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 SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+
+      - name: Upload binaries
+        if: matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-cublas-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
+
   emscripten:
     runs-on: ubuntu-latest
 
@@ -245,23 +340,113 @@ 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
+
+  ios:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+
+      - name: Configure
+        run: |
+          cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
+          mkdir models/ggml-base.en-encoder.mlmodelc
+
+      - 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
+
+  android:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        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
+
+      - name: Build
+        run: |
+          cd examples/whisper.android
+          ./gradlew assembleRelease --no-daemon
+
+  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

.github/workflows/examples.yml

@@ -0,0 +1,48 @@
+name: Examples Tests
+on:
+  push:
+    paths:
+      - examples/addon.node/**
+      - whisper.h
+  pull_request:
+    paths:
+      - examples/addon.node/**
+      - whisper.h
+
+jobs:
+  addon_node-ubuntu-latest:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        node-version: [ 16.x, 18.x ]
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install cmake
+          sudo apt-get install libsdl2-dev
+
+      - name: Use Node.js ${{ matrix.node-version }}
+        uses: actions/setup-node@v1
+        with:
+          node-version: ${{ matrix.node-version }}
+          cache: 'npm'
+
+      - name: Install package.json dependencies
+        working-directory: ./examples/addon.node
+        run: npm install
+
+      - name: Compile addon.node
+        run: npx cmake-js compile -T whisper-addon -B Release
+
+      - name: Download test model
+        run: |
+          bash ./models/download-ggml-model.sh base.en
+      - name: Test
+        run: |
+          cd examples/addon.node
+          npm run test

.gitignore

@@ -1,6 +1,8 @@
 *.o
 *.a
 .cache/
+.coreml/
+.test/
 .vs/
 .vscode/
 .DS_Store
@@ -9,7 +11,10 @@ build/
 build-em/
 build-debug/
 build-release/
+build-rwdi/
 build-static/
+build-cublas/
+build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
@@ -17,7 +22,10 @@ build-sanitize-thread/
 /stream
 /command
 /talk
+/talk-llama
 /bench
+/quantize
+/lsp
 
 arm_neon.h
 sync.sh
@@ -31,3 +39,10 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata
 
 extra/bench-gg.txt
+
+models/*.mlmodel
+models/*.mlmodelc
+models/*.mlpackage
+bindings/java/.gradle/
+bindings/java/.idea/
+.idea/

CMakeLists.txt

@@ -1,6 +1,6 @@
-cmake_minimum_required (VERSION 3.0)
+cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.2.0)
+project(whisper.cpp VERSION 1.4.2)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -35,30 +35,49 @@ endif()
 
 # options
 
-option(BUILD_SHARED_LIBS               "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
-
-option(WHISPER_ALL_WARNINGS            "whisper: enable all compiler warnings"                   ON)
-option(WHISPER_ALL_WARNINGS_3RD_PARTY  "whisper: enable all compiler warnings in 3rd party libs" OFF)
-
-option(WHISPER_SANITIZE_THREAD         "whisper: enable thread sanitizer"    OFF)
-option(WHISPER_SANITIZE_ADDRESS        "whisper: enable address sanitizer"   OFF)
-option(WHISPER_SANITIZE_UNDEFINED      "whisper: enable undefined sanitizer" OFF)
-
-option(WHISPER_BUILD_TESTS             "whisper: build tests"    ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_EXAMPLES          "whisper: build examples" ${WHISPER_STANDALONE})
-
-option(WHISPER_SUPPORT_SDL2            "whisper: support for libSDL2" OFF)
-
 if (APPLE)
-    option(WHISPER_NO_ACCELERATE       "whisper: disable Accelerate framework" OFF)
-    option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
-    option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
-    option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
+    set(WHISPER_METAL_DEFAULT ON)
 else()
-    option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
+    set(WHISPER_METAL_DEFAULT OFF)
 endif()
 
-option(WHISPER_PERF                    "whisper: enable perf timings" OFF)
+option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
+
+option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
+option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
+
+option(WHISPER_SANITIZE_THREAD        "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS       "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED     "whisper: enable undefined sanitizer" OFF)
+
+option(WHISPER_BUILD_TESTS            "whisper: build tests"    ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDALONE})
+
+option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
+
+option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
+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_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)
 
 # sanitizers
 
@@ -86,33 +105,179 @@ endif()
 
 find_package(Threads REQUIRED)
 
-# on APPLE - include Accelerate framework
-if (APPLE AND NOT WHISPER_NO_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
+# on APPLE
+if (APPLE)
+    # include Accelerate framework
+    if (NOT WHISPER_NO_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
 
-        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")
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            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")
+        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(WARNING "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)
+
+        if (COREML_FRAMEWORK)
+            message(STATUS "CoreML framework found")
+
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+        else()
+            message(WARNING "CoreML framework not found")
+        endif()
+
+        if (WHISPER_COREML_ALLOW_FALLBACK)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
+        endif()
     endif()
 endif()
 
-if (WHISPER_SUPPORT_OPENBLAS)
-    find_library(OPENBLAS_LIB
-        NAMES openblas libopenblas
-        )
-    if (OPENBLAS_LIB)
-        message(STATUS "OpenBLAS found")
+if (WHISPER_OPENBLAS)
+    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(WARNING "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_PREFER_PKGCONFIG 1)
+        set(BLA_SIZEOF_INTEGER 8)
+        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(WARNING "BLAS library was not found")
+        endif()
+    endif ()
+endif ()
+
+if (WHISPER_CUBLAS)
+    cmake_minimum_required(VERSION 3.17)
+
+    find_package(CUDAToolkit)
+
+    if (CUDAToolkit_FOUND)
+        message(STATUS "cuBLAS found")
+
+        enable_language(CUDA)
+
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+
+        add_compile_definitions(GGML_USE_CUBLAS)
+
+        if (WHISPER_STATIC)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        else()
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        endif()
 
-        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")
+        message(WARNING "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(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    endif()
+endif()
+
+if (WHISPER_CLBLAST)
+    find_package(CLBlast)
+    if (CLBlast_FOUND)
+        message(STATUS "CLBlast found")
+
+        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")
+    endif()
+endif()
+
+if( WHISPER_OPENVINO )
+    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+endif()
+
 # compiler flags
 
 if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
@@ -152,12 +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)
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
-        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
-        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
+        set(CMAKE_CXX_FLAGS "${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   "${CMAKE_C_FLAGS}   -pthread")
@@ -172,15 +350,109 @@ else()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
         endif()
     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()
 
+#
+# whisper.coreml - Core ML support
+#
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        )
+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
 #
@@ -190,6 +462,11 @@ set(TARGET whisper)
 add_library(${TARGET}
     ggml.h
     ggml.c
+    ggml-alloc.h
+    ggml-alloc.c
+    ${GGML_SOURCES_METAL}
+    ${GGML_SOURCES_CUDA}
+    ${GGML_SOURCES_OPENCL}
     whisper.h
     whisper.cpp
     )
@@ -200,6 +477,14 @@ target_include_directories(${TARGET} PUBLIC
     .
     )
 
+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})
 
@@ -215,7 +500,25 @@ if (BUILD_SHARED_LIBS)
 
     target_compile_definitions(${TARGET} PUBLIC
         WHISPER_SHARED
+        GGML_SHARED
         )
+
+    target_compile_definitions(${TARGET} PRIVATE
+        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_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")
 endif()
 
 if (EMSCRIPTEN)
@@ -226,10 +529,16 @@ target_compile_definitions(${TARGET} PUBLIC
     ${WHISPER_EXTRA_FLAGS}
     )
 
+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
     )
 
 #
@@ -242,7 +551,7 @@ add_subdirectory(bindings)
 # programs, examples and tests
 #
 
-if (WHISPER_BUILD_TESTS)
+if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     enable_testing()
     add_subdirectory(tests)
 endif ()

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2022 Georgi Gerganov
+Copyright (c) 2023 Georgi Gerganov
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -1,3 +1,5 @@
+default: main bench quantize
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -10,7 +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
@@ -30,25 +38,63 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# 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
@@ -56,65 +102,58 @@ 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)
-		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		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
-		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)
-		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
+		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 := $(shell sysinfo -cpu | grep "AVX2 ")
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
+
+		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
+		ifneq (,$(AVX2_M))
+			CFLAGS   += -mavx2
+			CXXFLAGS += -mavx2
 		endif
-		FMA_M := $(shell sysinfo -cpu | grep "FMA ")
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
+
+		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
+		ifneq (,$(FMA_M))
+			CFLAGS   += -mfma
+			CXXFLAGS += -mfma
 		endif
-		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
+
+		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
-	else
-		CFLAGS += -mfma -mf16c -mavx -mavx2
 	endif
 endif
-ifeq ($(UNAME_M),amd64)
-	CFLAGS += -mavx -mavx2 -mfma -mf16c
-endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -125,6 +164,7 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
 	endif
 endif
+
 ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
@@ -132,27 +172,107 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
+
+ifdef WHISPER_COREML
+	CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS  += -framework Foundation -framework CoreML
+
+ifdef WHISPER_COREML_ALLOW_FALLBACK
+	CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
+endif
+endif
+
+ifndef WHISPER_NO_METAL
+	ifeq ($(UNAME_S),Darwin)
+		WHISPER_METAL := 1
+
+		CXXFLAGS += -DGGML_USE_METAL
+		LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	endif
+endif
+
 ifdef WHISPER_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas
 endif
+
+ifdef WHISPER_CUBLAS
+	ifeq ($(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/$(UNAME_M)-linux/lib
+	WHISPER_OBJ += ggml-cuda.o
+	NVCC        = nvcc
+	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(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 $@ $<
+endif
+
+ifdef WHISPER_CLBLAST
+	CFLAGS 		+= -DGGML_USE_CLBLAST
+	CXXFLAGS 	+= -DGGML_USE_CLBLAST
+	LDFLAGS	 	+= -lclblast
+	ifeq ($(UNAME_S),Darwin)
+		LDFLAGS	 	+= -framework OpenCL
+	else
+		LDFLAGS	    += -lOpenCL
+	endif
+	WHISPER_OBJ	+= ggml-opencl.o
+
+ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 ifdef WHISPER_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
 endif
+
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS   += -mcpu=native
+	CXXFLAGS += -mcpu=native
 endif
+
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, 2, 3
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	# 32-bit Raspberry Pi 1, 2, 3
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
 endif
+
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 4
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# 32-bit ARM, for example on Armbian or possibly raspbian
+	#CFLAGS   += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	#CXXFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+
+	# 64-bit ARM on 32-bit OS, use these (TODO: auto-detect 64-bit)
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
+
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
 
 #
@@ -170,26 +290,48 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main
-
 #
 # Build library
 #
 
-ggml.o: ggml.c ggml.h
-	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
+ggml.o: ggml.c ggml.h ggml-cuda.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-whisper.o: whisper.cpp whisper.h
-	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
+ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-libwhisper.a: ggml.o whisper.o
-	$(AR) rcs libwhisper.a ggml.o whisper.o
+WHISPER_OBJ += ggml-alloc.o
 
-libwhisper.so: ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
+whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+ifndef WHISPER_COREML
+WHISPER_OBJ += whisper.o
+else
+whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder.mm -o whisper-encoder.o
+
+whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-impl.h
+	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
+
+WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
+endif
+
+ifdef WHISPER_METAL
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
+
+WHISPER_OBJ += ggml-metal.o
+endif
+
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize lsp libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -197,21 +339,33 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-main: examples/main/main.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o whisper.o -o main $(LDFLAGS)
+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 -h
 
-stream: examples/stream/stream.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-command: examples/command/command.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp ggml.o whisper.o -o command $(CC_SDL) $(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)
 
-talk: examples/talk/talk.cpp  examples/talk/gpt-2.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp ggml.o whisper.o -o talk $(CC_SDL) $(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)
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(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)
+
+lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o lsp $(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)
+
+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)
 
 #
 # Audio samples
@@ -226,12 +380,19 @@ 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
@@ -273,4 +434,4 @@ 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))

README.md

@@ -1,21 +1,28 @@
 # whisper.cpp
 
+![whisper.cpp](https://user-images.githubusercontent.com/1991296/235238348-05d0f6a4-da44-4900-a1de-d0707e75b763.jpeg)
+
 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![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/)
 
-Stable: [v1.2.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+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)
 
 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 and Accelerate framework
+- 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
+- Support for CPU-only inference
+- [Partial GPU support for NVIDIA via cuBLAS](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)
+- [BLAS CPU support via OpenBLAS](https://github.com/ggerganov/whisper.cpp#blas-cpu-support-via-openblas)
+- [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:
@@ -23,6 +30,7 @@ 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)]
@@ -42,6 +50,10 @@ 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 Apply 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
@@ -53,17 +65,21 @@ 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
 
-First, download one of the Whisper models converted in [ggml format](models). For example:
+First clone the repository.
+
+Then, download one of the Whisper models converted in [ggml format](models). For example:
 
 ```bash
 bash ./models/download-ggml-model.sh base.en
 ```
 
+If you wish to convert the Whisper models to ggml format yourself, instructions are in [models/README.md](models/README.md).
+
 Now build the [main](examples/main) example and transcribe an audio file like this:
 
 ```bash
@@ -104,6 +120,7 @@ options:
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
   -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
   -tr,       --translate         [false  ] translate from source language to english
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
   -di,       --diarize           [false  ] stereo audio diarization
   -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
   -otxt,     --output-txt        [false  ] output result in a text file
@@ -223,10 +240,211 @@ make large
 | medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
 | large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
 
+## Quantization
+
+`whisper.cpp` supports integer quantization of the Whisper `ggml` models.
+Quantized models require less memory and disk space and depending on the hardware can be processed more efficiently.
+
+Here are the steps for creating and using a quantized model:
+
+```bash
+# quantize a model with Q5_0 method
+make quantize
+./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
+
+# run the examples as usual, specifying the quantized model file
+./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
+```
+
+## Core ML support
+
+On Apple Silicon devices, the Encoder inference can be executed on the Apple Neural Engine (ANE) via Core ML. This can result in significant
+speed-up - more than x3 faster compared with CPU-only execution. Here are the instructions for generating a Core ML model and using it with `whisper.cpp`:
+
+- Install Python dependencies needed for the creation of the Core ML model:
+
+  ```bash
+  pip install ane_transformers
+  pip install openai-whisper
+  pip install coremltools
+  ```
+
+  - To ensure `coremltools` operates correctly, please confirm that [Xcode](https://developer.apple.com/xcode/) is installed and execute `xcode-select --install` to install the command-line tools.
+  - Python 3.10 is recommended.
+  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html)  for this step:
+    - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
+    - To activate the environment, use: `conda activate py310-whisper`
+
+- Generate a Core ML model. For example, to generate a `base.en` model, use:
+
+  ```bash
+  ./models/generate-coreml-model.sh base.en
+  ```
+
+  This will generate the folder `models/ggml-base.en-encoder.mlmodelc`
+
+- Build `whisper.cpp` with Core ML support:
+
+  ```bash
+  # using Makefile
+  make clean
+  WHISPER_COREML=1 make -j
+
+  # using CMake
+  cmake -B build -DWHISPER_COREML=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_init_state: loading Core ML model from 'models/ggml-base.en-encoder.mlmodelc'
+  whisper_init_state: first run on a device may take a while ...
+  whisper_init_state: Core ML model loaded
+
+  system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | COREML = 1 |
+
+  ...
+  ```
+
+  The first run on a device is slow, since the ANE service compiles the Core ML model to some device-specific format.
+  Next runs are faster.
+
+For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
+
+## OpenVINO support
+
+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 via cuBLAS
+
+With NVIDIA cards the Encoder processing can to a large extent be offloaded to the GPU through cuBLAS.
+First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads
+
+Now build `whisper.cpp` with cuBLAS support:
+
+```
+make clean
+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 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
+
+Now build `whisper.cpp` with CLBlast support:
+
+```
+Makefile:
+cd whisper.cpp
+make clean
+WHISPER_CLBLAST=1 make -j
+
+CMake:
+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
-- No GPU support (yet)
 
 ## Another example
 
@@ -313,7 +531,7 @@ whisper_print_timings:    total time = 32733.52 ms
 ## Real-time audio input example
 
 This is a naive example of performing real-time inference on audio from your microphone.
-The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continously.
+The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```java
@@ -328,6 +546,10 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:
 
+```java
+./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
+```
+
 <img width="965" alt="image" src="https://user-images.githubusercontent.com/1991296/197356445-311c8643-9397-4e5e-b46e-0b4b4daa2530.png">
 
 ## Controlling the length of the generated text segments (experimental)
@@ -354,7 +576,7 @@ 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`:
 
@@ -367,7 +589,7 @@ system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
-[00:00:00.000 --> 00:00:00.320]  
+[00:00:00.000 --> 00:00:00.320]
 [00:00:00.320 --> 00:00:00.370]   And
 [00:00:00.370 --> 00:00:00.690]   so
 [00:00:00.690 --> 00:00:00.850]   my
@@ -395,6 +617,32 @@ 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
@@ -433,6 +681,19 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ---
 
+## Video comparison of different models
+
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
+
+```java
+./extra/bench-wts.sh samples/jfk.wav
+ffplay ./samples/jfk.wav.all.mp4
+```
+
+https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8c3b-eb9f9c39d6fc.mp4
+
+---
+
 ## Benchmarks
 
 In order to have an objective comparison of the performance of the inference across different system configurations,
@@ -453,7 +714,7 @@ The original models are converted to a custom binary format. This allows to pack
 You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
 or manually from here:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
@@ -463,9 +724,21 @@ in [models](models).
 
 - [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)
+  - 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)
-- [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
+- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+  - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
+  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
+- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
@@ -479,6 +752,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
 | [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |

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

@@ -17,9 +17,9 @@ import (
 // CONSTANTS
 
 const (
-	srcUrl  = "https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main" // The location of the models
-	srcExt  = ".bin"                                                               // Filename extension
-	bufSize = 1024 * 64                                                            // Size of the buffer used for downloading the model
+	srcUrl  = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main" // The location of the models
+	srcExt  = ".bin"                                                      // Filename extension
+	bufSize = 1024 * 64                                                   // Size of the buffer used for downloading the model
 )
 
 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)
 }
@@ -105,6 +109,10 @@ func (p *Params) SetMaxSegmentLength(n int) {
 	p.max_len = C.int(n)
 }
 
+func (p *Params) SetTokenTimestamps(b bool) {
+	p.token_timestamps = toBool(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)

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)
@@ -111,6 +115,11 @@ func (context *context) SetMaxSegmentLength(n uint) {
 	context.params.SetMaxSegmentLength(int(n))
 }
 
+// Set token timestamps flag
+func (context *context) SetTokenTimestamps(b bool) {
+	context.params.SetTokenTimestamps(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
@@ -147,12 +156,16 @@ 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)
 	}
 
@@ -160,24 +173,28 @@ 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
 	}
@@ -280,10 +297,14 @@ func toSegment(ctx *whisper.Context, n int) Segment {
 func toTokens(ctx *whisper.Context, n int) []Token {
 	result := make([]Token, ctx.Whisper_full_n_tokens(n))
 	for i := 0; i < len(result); i++ {
+		data := ctx.Whisper_full_get_token_data(n, i)
+
 		result[i] = Token{
-			Id:   int(ctx.Whisper_full_get_token_id(n, i)),
-			Text: strings.TrimSpace(ctx.Whisper_full_get_token_text(n, i)),
-			P:    ctx.Whisper_full_get_token_p(n, i),
+			Id:    int(ctx.Whisper_full_get_token_id(n, i)),
+			Text:  ctx.Whisper_full_get_token_text(n, i),
+			P:     ctx.Whisper_full_get_token_p(n, i),
+			Start: time.Duration(data.T0()) * time.Millisecond * 10,
+			End:   time.Duration(data.T1()) * time.Millisecond * 10,
 		}
 	}
 	return result

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 {
@@ -38,15 +42,17 @@ 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)
 
 	// 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.
@@ -85,7 +91,8 @@ type Segment struct {
 
 // Token is a text or special token
 type Token struct {
-	Id   int
-	Text string
-	P    float32
+	Id         int
+	Text       string
+	P          float32
+	Start, End time.Duration
 }

bindings/go/pkg/whisper/model.go

@@ -94,6 +94,7 @@ func (model *model) NewContext() (Context, error) {
 	params.SetPrintRealtime(false)
 	params.SetPrintTimestamps(false)
 	params.SetThreads(runtime.NumCPU())
+	params.SetNoContext(true)
 
 	// Return new context
 	return newContext(model, params)

bindings/go/whisper.go

@@ -15,21 +15,31 @@ 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
 // Called on every newly generated text segment
 // Use the whisper_full_...() functions to obtain the text segments
-static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void* user_data) {
+static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         callNewSegment(user_data, n_new);
     }
 }
 
+// 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
-static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, struct whisper_state* state, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         return callEncoderBegin(user_data);
     }
@@ -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;
 }
 */
@@ -258,13 +270,13 @@ 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 +302,19 @@ 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 +338,18 @@ 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 {
@@ -356,7 +388,7 @@ func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {
 
 // Get token data for the specified token in the specified segment.
 // This contains probabilities, timestamps, etc.
-func (ctx *Context) whisper_full_get_token_data(segment int, token int) TokenData {
+func (ctx *Context) Whisper_full_get_token_data(segment int, token int) TokenData {
 	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
 }
 
@@ -370,6 +402,7 @@ 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 +414,14 @@ 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 +437,13 @@ 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 {
@@ -407,3 +455,15 @@ func callEncoderBegin(user_data unsafe.Pointer) C.bool {
 	}
 	return true
 }
+
+func (t TokenData) T0() int64 {
+	return int64(t.t0)
+}
+
+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 +1,132 @@
+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 +1,39 @@
+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 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() */
+    String path_model;
+
+//    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 +1,151 @@
+package io.github.ggerganov.whispercpp;
+
+import com.sun.jna.Native;
+import com.sun.jna.Pointer;
+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;
+
+/**
+ * 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 greedyPointer = null;
+    private Pointer beamPointer = 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 {
+        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(modelPath);
+
+        if (ctx == null) {
+            throw new FileNotFoundException(modelPath);
+        }
+    }
+
+    /**
+     * 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 (greedyPointer == null) {
+                greedyPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            }
+            pointer = greedyPointer;
+        } else {
+            if (beamPointer == null) {
+                beamPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            }
+            pointer = beamPointer;
+        }
+
+        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 (greedyPointer != null) {
+            Native.free(Pointer.nativeValue(greedyPointer));
+            greedyPointer = null;
+        }
+        if (beamPointer != null) {
+            Native.free(Pointer.nativeValue(beamPointer));
+            beamPointer = 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 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 +1,376 @@
+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.WhisperFullParams;
+
+public interface WhisperCppJnaLibrary extends Library {
+    WhisperCppJnaLibrary instance = Native.load("whisper", WhisperCppJnaLibrary.class);
+
+    String whisper_print_system_info();
+
+    /**
+     * 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);
+
+    /**
+     * 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/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/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 +1,321 @@
+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;
+    }
+
+    /** 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);
+    }
+
+    @Override
+    protected List<String> getFieldOrder() {
+        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
+                "no_context", "single_segment",
+                "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");
+    }
+}

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 +1,102 @@
+package io.github.ggerganov.whispercpp;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+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;
+
+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.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(2, 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(2, 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();
+        }
+    }
+}

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/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": "1.2.0",
+  "version": "1.4.2",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/.gitignore

@@ -0,0 +1,7 @@
+Makefile
+ggml.c
+ggml.h
+whisper.bundle
+whisper.cpp
+whisper.h
+dr_wav.h

bindings/ruby/ext/extconf.rb

@@ -0,0 +1,21 @@
+require 'mkmf'
+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__),'..','..','..','examples','dr_wav.h')} .")
+
+
+# need to use c++ compiler flags
+$CXXFLAGS << ' -std=c++11'
+# Set to true when building binary gems
+if enable_config('static-stdlib', false)
+  $LDFLAGS << ' -static-libgcc -static-libstdc++'
+end
+
+if enable_config('march-tune-native', false)
+  $CFLAGS << ' -march=native -mtune=native'
+  $CXXFLAGS << ' -march=native -mtune=native'
+end
+
+create_makefile('whisper')

bindings/ruby/ext/ruby_whisper.cpp

@@ -0,0 +1,426 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BOOL_PARAMS_SETTER(self, prop, value) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (value == Qfalse || value == Qnil) { \
+    rwp->params.prop = false; \
+  } else { \
+    rwp->params.prop = true; \
+  } \
+  return value; \
+
+#define BOOL_PARAMS_GETTER(self,  prop) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (rwp->params.prop) { \
+    return Qtrue; \
+  } else { \
+    return Qfalse; \
+  }
+
+VALUE mWhisper;
+VALUE cContext;
+VALUE cParams;
+
+static void ruby_whisper_free(ruby_whisper *rw) {
+  if (rw->context) {
+    whisper_free(rw->context);
+    rw->context = NULL;
+  }
+}
+static void ruby_whisper_params_free(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_mark(ruby_whisper *rw) {
+  // call rb_gc_mark on any ruby references in rw
+}
+
+void rb_whisper_free(ruby_whisper *rw) {
+  ruby_whisper_free(rw);
+  free(rw);
+}
+
+void rb_whisper_params_mark(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_params_free(ruby_whisper_params *rwp) {
+  ruby_whisper_params_free(rwp);
+  free(rwp);
+}
+
+static VALUE ruby_whisper_allocate(VALUE klass) {
+  ruby_whisper *rw;
+  rw = ALLOC(ruby_whisper);
+  rw->context = NULL;
+  return Data_Wrap_Struct(klass, rb_whisper_mark, rb_whisper_free, rw);
+}
+
+static VALUE ruby_whisper_params_allocate(VALUE klass) {
+  ruby_whisper_params *rwp;
+  rwp = ALLOC(ruby_whisper_params);
+  rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  return Data_Wrap_Struct(klass, rb_whisper_params_mark, rb_whisper_params_free, rwp);
+}
+
+static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  VALUE whisper_model_file_path;
+
+  // TODO: we can support init from buffer here too maybe another ruby object to expose
+  rb_scan_args(argc, argv, "01", &whisper_model_file_path);
+  Data_Get_Struct(self, ruby_whisper, rw);
+
+  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));
+  if (rw->context == nullptr) {
+    rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
+  }
+  return self;
+}
+
+/*
+ * transcribe a single file
+ * can emit to a block results
+ *
+ **/
+static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  VALUE wave_file_path, blk, params;
+
+  rb_scan_args(argc, argv, "02&", &wave_file_path, &params, &blk);
+  Data_Get_Struct(self, ruby_whisper, rw);
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+
+  if (!rb_respond_to(wave_file_path, rb_intern("to_s"))) {
+    rb_raise(rb_eRuntimeError, "Expected file path to wave file");
+  }
+
+  std::string fname_inp = StringValueCStr(wave_file_path);
+
+  std::vector<float> pcmf32; // mono-channel F32 PCM
+  std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+  // WAV input - this is directly from main.cpp example
+  {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname_inp == "-") {
+      {
+        uint8_t buf[1024];
+        while (true) {
+          const size_t n = fread(buf, 1, sizeof(buf), stdin);
+          if (n == 0) {
+            break;
+          }
+          wav_data.insert(wav_data.end(), buf, buf + n);
+        }
+      }
+
+      if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open WAV file from stdin\n");
+        return self;
+      }
+
+      fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    } else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
+      fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+      fprintf(stderr, "WAV file '%s' must be mono or stereo\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (rwp->diarize && wav.channels != 2 && rwp->params.print_timestamps == false) {
+      fprintf(stderr, "WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
+      fprintf(stderr, "WAV file '%s' must be %i kHz\n", fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
+      return self;
+    }
+
+    if (wav.bitsPerSample != 16) {
+      fprintf(stderr, "WAV file '%s' must be 16-bit\n", fname_inp.c_str());
+      return self;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[i])/32768.0f;
+      }
+    } else {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+      }
+    }
+
+    if (rwp->diarize) {
+      // convert to stereo, float
+      pcmf32s.resize(2);
+
+      pcmf32s[0].resize(n);
+      pcmf32s[1].resize(n);
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+        pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+      }
+    }
+  }
+  {
+    static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+      bool is_aborted = *(bool*)user_data;
+      return !is_aborted;
+    };
+    rwp->params.encoder_begin_callback_user_data = &is_aborted;
+  }
+
+  if (whisper_full_parallel(rw->context, rwp->params, pcmf32.data(), pcmf32.size(), 1) != 0) {
+    fprintf(stderr, "failed to process audio\n");
+    return self;
+  }
+  const int n_segments = whisper_full_n_segments(rw->context);
+  VALUE output = rb_str_new2("");
+  for (int i = 0; i < n_segments; ++i) {
+    const char * text = whisper_full_get_segment_text(rw->context, i);
+    output = rb_str_concat(output, rb_str_new2(text));
+  }
+  VALUE idCall = rb_intern("call");
+  if (blk != Qnil) {
+    rb_funcall(blk, idCall, 1, output);
+  }
+  return self;
+}
+
+/*
+ * params.language = "auto" | "en", etc...
+ */
+static VALUE ruby_whisper_params_set_language(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (value == Qfalse || value == Qnil) {
+    rwp->params.language = "auto";
+  } else {
+    rwp->params.language = StringValueCStr(value);
+  }
+  return value;
+}
+static VALUE ruby_whisper_params_get_language(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (rwp->params.language) {
+    return rb_str_new2(rwp->params.language);
+  } else {
+    return rb_str_new2("auto");
+  }
+}
+static VALUE ruby_whisper_params_set_translate(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, translate, value)
+}
+static VALUE ruby_whisper_params_get_translate(VALUE self) {
+  BOOL_PARAMS_GETTER(self, translate)
+}
+static VALUE ruby_whisper_params_set_no_context(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, no_context, value)
+}
+static VALUE ruby_whisper_params_get_no_context(VALUE self) {
+  BOOL_PARAMS_GETTER(self, no_context)
+}
+static VALUE ruby_whisper_params_set_single_segment(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, single_segment, value)
+}
+static VALUE ruby_whisper_params_get_single_segment(VALUE self) {
+  BOOL_PARAMS_GETTER(self, single_segment)
+}
+static VALUE ruby_whisper_params_set_print_special(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_special, value)
+}
+static VALUE ruby_whisper_params_get_print_special(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_special)
+}
+static VALUE ruby_whisper_params_set_print_progress(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_progress, value)
+}
+static VALUE ruby_whisper_params_get_print_progress(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_progress)
+}
+static VALUE ruby_whisper_params_set_print_realtime(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_realtime, value)
+}
+static VALUE ruby_whisper_params_get_print_realtime(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_realtime)
+}
+static VALUE ruby_whisper_params_set_print_timestamps(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, print_timestamps, value)
+}
+static VALUE ruby_whisper_params_get_print_timestamps(VALUE self) {
+  BOOL_PARAMS_GETTER(self, print_timestamps)
+}
+static VALUE ruby_whisper_params_set_suppress_blank(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, suppress_blank, value)
+}
+static VALUE ruby_whisper_params_get_suppress_blank(VALUE self) {
+  BOOL_PARAMS_GETTER(self, suppress_blank)
+}
+static VALUE ruby_whisper_params_set_suppress_non_speech_tokens(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, suppress_non_speech_tokens, value)
+}
+static VALUE ruby_whisper_params_get_suppress_non_speech_tokens(VALUE self) {
+  BOOL_PARAMS_GETTER(self, suppress_non_speech_tokens)
+}
+static VALUE ruby_whisper_params_get_token_timestamps(VALUE self) {
+  BOOL_PARAMS_GETTER(self, token_timestamps)
+}
+static VALUE ruby_whisper_params_set_token_timestamps(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, token_timestamps, value)
+}
+static VALUE ruby_whisper_params_get_split_on_word(VALUE self) {
+  BOOL_PARAMS_GETTER(self, split_on_word)
+}
+static VALUE ruby_whisper_params_set_split_on_word(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, split_on_word, value)
+}
+static VALUE ruby_whisper_params_get_speed_up(VALUE self) {
+  BOOL_PARAMS_GETTER(self, speed_up)
+}
+static VALUE ruby_whisper_params_set_speed_up(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, speed_up, value)
+}
+static VALUE ruby_whisper_params_get_diarize(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (rwp->diarize) {
+    return Qtrue;
+  } else {
+    return Qfalse;
+  }
+}
+static VALUE ruby_whisper_params_set_diarize(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  if (value == Qfalse || value == Qnil) {
+    rwp->diarize = false;
+  } else {
+    rwp->diarize = true;
+  } \
+  return value;
+}
+
+static VALUE ruby_whisper_params_get_offset(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.offset_ms);
+}
+static VALUE ruby_whisper_params_set_offset(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.offset_ms = NUM2INT(value);
+  return value;
+}
+static VALUE ruby_whisper_params_get_duration(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.duration_ms);
+}
+static VALUE ruby_whisper_params_set_duration(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.duration_ms = NUM2INT(value);
+  return value;
+}
+
+static VALUE ruby_whisper_params_get_max_text_tokens(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return INT2NUM(rwp->params.n_max_text_ctx);
+}
+static VALUE ruby_whisper_params_set_max_text_tokens(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.n_max_text_ctx = NUM2INT(value);
+  return value;
+}
+
+void Init_whisper() {
+  mWhisper = rb_define_module("Whisper");
+  cContext = rb_define_class_under(mWhisper, "Context", rb_cObject);
+  cParams  = rb_define_class_under(mWhisper, "Params", rb_cObject);
+
+  rb_define_alloc_func(cContext, ruby_whisper_allocate);
+  rb_define_method(cContext, "initialize", ruby_whisper_initialize, -1);
+
+  rb_define_method(cContext, "transcribe", ruby_whisper_transcribe, -1);
+
+  rb_define_alloc_func(cParams, ruby_whisper_params_allocate);
+
+  rb_define_method(cParams, "language=", ruby_whisper_params_set_language, 1);
+  rb_define_method(cParams, "language", ruby_whisper_params_get_language, 0);
+  rb_define_method(cParams, "translate=", ruby_whisper_params_set_translate, 1);
+  rb_define_method(cParams, "translate", ruby_whisper_params_get_translate, 0);
+  rb_define_method(cParams, "no_context=", ruby_whisper_params_set_no_context, 1);
+  rb_define_method(cParams, "no_context", ruby_whisper_params_get_no_context, 0);
+  rb_define_method(cParams, "single_segment=", ruby_whisper_params_set_single_segment, 1);
+  rb_define_method(cParams, "single_segment", ruby_whisper_params_get_single_segment, 0);
+  rb_define_method(cParams, "print_special", ruby_whisper_params_get_print_special, 0);
+  rb_define_method(cParams, "print_special=", ruby_whisper_params_set_print_special, 1);
+  rb_define_method(cParams, "print_progress", ruby_whisper_params_get_print_progress, 0);
+  rb_define_method(cParams, "print_progress=", ruby_whisper_params_set_print_progress, 1);
+  rb_define_method(cParams, "print_realtime", ruby_whisper_params_get_print_realtime, 0);
+  rb_define_method(cParams, "print_realtime=", ruby_whisper_params_set_print_realtime, 1);
+  rb_define_method(cParams, "print_timestamps", ruby_whisper_params_get_print_timestamps, 0);
+  rb_define_method(cParams, "print_timestamps=", ruby_whisper_params_set_print_timestamps, 1);
+  rb_define_method(cParams, "suppress_blank", ruby_whisper_params_get_suppress_blank, 0);
+  rb_define_method(cParams, "suppress_blank=", ruby_whisper_params_set_suppress_blank, 1);
+  rb_define_method(cParams, "suppress_non_speech_tokens", ruby_whisper_params_get_suppress_non_speech_tokens, 0);
+  rb_define_method(cParams, "suppress_non_speech_tokens=", ruby_whisper_params_set_suppress_non_speech_tokens, 1);
+  rb_define_method(cParams, "token_timestamps", ruby_whisper_params_get_token_timestamps, 0);
+  rb_define_method(cParams, "token_timestamps=", ruby_whisper_params_set_token_timestamps, 1);
+  rb_define_method(cParams, "split_on_word", ruby_whisper_params_get_split_on_word, 0);
+  rb_define_method(cParams, "split_on_word=", ruby_whisper_params_set_split_on_word, 1);
+  rb_define_method(cParams, "speed_up", ruby_whisper_params_get_speed_up, 0);
+  rb_define_method(cParams, "speed_up=", ruby_whisper_params_set_speed_up, 1);
+  rb_define_method(cParams, "diarize", ruby_whisper_params_get_diarize, 0);
+  rb_define_method(cParams, "diarize=", ruby_whisper_params_set_diarize, 1);
+
+  rb_define_method(cParams, "offset", ruby_whisper_params_get_offset, 0);
+  rb_define_method(cParams, "offset=", ruby_whisper_params_set_offset, 1);
+  rb_define_method(cParams, "duration", ruby_whisper_params_get_duration, 0);
+  rb_define_method(cParams, "duration=", ruby_whisper_params_set_duration, 1);
+
+  rb_define_method(cParams, "max_text_tokens", ruby_whisper_params_get_max_text_tokens, 0);
+  rb_define_method(cParams, "max_text_tokens=", ruby_whisper_params_set_max_text_tokens, 1);
+}
+#ifdef __cplusplus
+}
+#endif

bindings/ruby/ext/ruby_whisper.h

@@ -0,0 +1,15 @@
+#ifndef __RUBY_WHISPER_H
+#define __RUBY_WHISPER_H
+
+#include "whisper.h"
+
+typedef struct {
+  struct whisper_context *context;
+} ruby_whisper;
+
+typedef struct {
+  struct whisper_full_params params;
+  bool diarize;
+} ruby_whisper_params;
+
+#endif

bindings/ruby/tests/test_whisper.rb

@@ -0,0 +1,138 @@
+TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+EXTDIR = File.join(TOPDIR, 'ext')
+#$LIBDIR = File.join(TOPDIR, 'lib')
+#$:.unshift(LIBDIR)
+$:.unshift(EXTDIR)
+
+require 'whisper'
+require 'test/unit'
+
+class TestWhisper < Test::Unit::TestCase
+  def setup
+    @params  = Whisper::Params.new
+  end
+
+  def test_language
+    @params.language = "en"
+    assert_equal @params.language, "en"
+    @params.language = "auto"
+    assert_equal @params.language, "auto"
+  end
+
+  def test_offset
+    @params.offset = 10_000
+    assert_equal @params.offset, 10_000
+    @params.offset = 0
+    assert_equal @params.offset, 0
+  end
+
+  def test_duration
+    @params.duration = 60_000
+    assert_equal @params.duration, 60_000
+    @params.duration = 0
+    assert_equal @params.duration, 0
+  end
+
+  def test_max_text_tokens
+    @params.max_text_tokens = 300
+    assert_equal @params.max_text_tokens, 300
+    @params.max_text_tokens = 0
+    assert_equal @params.max_text_tokens, 0
+  end
+
+  def test_translate
+    @params.translate = true
+    assert @params.translate
+    @params.translate = false
+    assert !@params.translate
+  end
+
+  def test_no_context
+    @params.no_context = true
+    assert @params.no_context
+    @params.no_context = false
+    assert !@params.no_context
+  end
+
+  def test_single_segment
+    @params.single_segment = true
+    assert @params.single_segment
+    @params.single_segment = false
+    assert !@params.single_segment
+  end
+
+  def test_print_special
+    @params.print_special = true
+    assert @params.print_special
+    @params.print_special = false
+    assert !@params.print_special
+  end
+
+  def test_print_progress
+    @params.print_progress = true
+    assert @params.print_progress
+    @params.print_progress = false
+    assert !@params.print_progress
+  end
+
+  def test_print_realtime
+    @params.print_realtime = true
+    assert @params.print_realtime
+    @params.print_realtime = false
+    assert !@params.print_realtime
+  end
+
+  def test_print_timestamps
+    @params.print_timestamps = true
+    assert @params.print_timestamps
+    @params.print_timestamps = false
+    assert !@params.print_timestamps
+  end
+
+  def test_suppress_blank
+    @params.suppress_blank = true
+    assert @params.suppress_blank
+    @params.suppress_blank = false
+    assert !@params.suppress_blank
+  end
+
+  def test_suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = true
+    assert @params.suppress_non_speech_tokens
+    @params.suppress_non_speech_tokens = false
+    assert !@params.suppress_non_speech_tokens
+  end
+
+  def test_token_timestamps
+    @params.token_timestamps = true
+    assert @params.token_timestamps
+    @params.token_timestamps = false
+    assert !@params.token_timestamps
+  end
+
+  def test_split_on_word
+    @params.split_on_word = true
+    assert @params.split_on_word
+    @params.split_on_word = false
+    assert !@params.split_on_word
+  end
+
+  def test_speed_up
+    @params.speed_up = true
+    assert @params.speed_up
+    @params.speed_up = false
+    assert !@params.speed_up
+  end
+
+  def test_whisper
+    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
+    params  = Whisper::Params.new
+    params.print_timestamps = false
+
+    jfk = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+    @whisper.transcribe(jfk, params) {|text|
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, text
+    }
+  end
+
+end

coreml/whisper-decoder-impl.h

@@ -0,0 +1,146 @@
+//
+// whisper-decoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implInput : NSObject<MLFeatureProvider>
+
+/// token_data as 1 by 1 matrix of 32-bit integers
+@property (readwrite, nonatomic, strong) MLMultiArray * token_data;
+
+/// audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * audio_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implOutput : NSObject<MLFeatureProvider>
+
+/// var_1346 as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * var_1346;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @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_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param options prediction options
+    @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_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param token_data as 1 by 1 matrix of 32-bit integers:
+    @param audio_data as 1 × 384 × 1 × 1500 4-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_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_decoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @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 predictions as NSArray<whisper_decoder_implOutput *>
+*/
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-decoder-impl.m

@@ -0,0 +1,201 @@
+//
+// whisper-decoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-decoder-impl.h"
+
+@implementation whisper_decoder_implInput
+
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data {
+    self = [super init];
+    if (self) {
+        _token_data = token_data;
+        _audio_data = audio_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"token_data", @"audio_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"token_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.token_data];
+    }
+    if ([featureName isEqualToString:@"audio_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.audio_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_implOutput
+
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 {
+    self = [super init];
+    if (self) {
+        _var_1346 = var_1346;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"var_1346"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"var_1346"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.var_1346];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_decoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-decoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_decoder_impl *typedModel = [[whisper_decoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[outFeatures featureValueForName:@"var_1346"].multiArrayValue];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_decoder_implInput *input_ = [[whisper_decoder_implInput alloc] initWithToken_data:token_data audio_data:audio_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_decoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[resultProvider featureValueForName:@"var_1346"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder-impl.h

@@ -0,0 +1,142 @@
+//
+// whisper-encoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implInput : NSObject<MLFeatureProvider>
+
+/// logmel_data as 1 × 80 × 3000 3-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * logmel_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implOutput : NSObject<MLFeatureProvider>
+
+/// output as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * output;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithOutput:(MLMultiArray *)output NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @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
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param options prediction options
+    @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
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param logmel_data as 1 × 80 × 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
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_encoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @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 predictions as NSArray<whisper_encoder_implOutput *>
+*/
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-encoder-impl.m

@@ -0,0 +1,197 @@
+//
+// whisper-encoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-encoder-impl.h"
+
+@implementation whisper_encoder_implInput
+
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data {
+    self = [super init];
+    if (self) {
+        _logmel_data = logmel_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"logmel_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"logmel_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.logmel_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_implOutput
+
+- (instancetype)initWithOutput:(MLMultiArray *)output {
+    self = [super init];
+    if (self) {
+        _output = output;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"output"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"output"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.output];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_encoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-encoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @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.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_encoder_impl *typedModel = [[whisper_encoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_encoder_implInput *input_ = [[whisper_encoder_implInput alloc] initWithLogmel_data:logmel_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_encoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_encoder_implOutput * result = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder.h

@@ -0,0 +1,22 @@
+// Wrapper of the Core ML Whisper Encoder model
+//
+// Code is derived from the work of Github user @wangchou
+// ref: https://github.com/wangchou/callCoreMLFromCpp
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context;
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model);
+void whisper_coreml_free(struct whisper_coreml_context * ctx);
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out);
+
+#if __cplusplus
+}
+#endif

coreml/whisper-encoder.mm

@@ -0,0 +1,71 @@
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-encoder.h"
+#import "whisper-encoder-impl.h"
+
+#import <CoreML/CoreML.h>
+
+#include <stdlib.h>
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context {
+    const void * data;
+};
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
+    NSString * path_model_str = [[NSString alloc] initWithUTF8String:path_model];
+
+    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
+
+    // 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;
+    }
+
+    whisper_coreml_context * ctx = new whisper_coreml_context;
+
+    ctx->data = data;
+
+    return ctx;
+}
+
+void whisper_coreml_free(struct whisper_coreml_context * ctx) {
+    CFRelease(ctx->data);
+    delete ctx;
+}
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out) {
+    MLMultiArray * inMultiArray = [
+        [MLMultiArray alloc] initWithDataPointer: mel
+                                           shape: @[@1, @80, @3000]
+                                        dataType: MLMultiArrayDataTypeFloat32
+                                         strides: @[@(240000), @(3000), @1]
+                                     deallocator: nil
+                                           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));
+    }
+}
+
+#if __cplusplus
+}
+#endif

examples/CMakeLists.txt

@@ -4,7 +4,7 @@ find_package(Threads REQUIRED)
 
 # third-party
 
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # SDL2
     find_package(SDL2 REQUIRED)
 
@@ -14,6 +14,41 @@ if (WHISPER_SUPPORT_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
+# common
+
+set(TARGET common)
+
+add_library(${TARGET} STATIC
+    common.h
+    common.cpp
+    common-ggml.h
+    common-ggml.cpp
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE whisper)
+
+set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+if (WHISPER_SDL2)
+    # common-sdl
+
+    set(TARGET common-sdl)
+
+    add_library(${TARGET} STATIC
+        common-sdl.h
+        common-sdl.cpp
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})
+
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -31,5 +66,8 @@ else()
     add_subdirectory(stream)
     add_subdirectory(command)
     add_subdirectory(bench)
+    add_subdirectory(quantize)
     add_subdirectory(talk)
+    add_subdirectory(talk-llama)
+    add_subdirectory(lsp)
 endif()

examples/addon.node/CMakeLists.txt

@@ -23,4 +23,9 @@ string(REPLACE "\"" "" NODE_ADDON_API_DIR ${NODE_ADDON_API_DIR})
 target_include_directories(${TARGET} PRIVATE ${NODE_ADDON_API_DIR})
 #==================================================================
 
-target_link_libraries(${TARGET} ${CMAKE_JS_LIB} whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} ${CMAKE_JS_LIB} common whisper ${CMAKE_THREAD_LIBS_INIT})
+
+if(MSVC AND CMAKE_JS_NODELIB_DEF AND CMAKE_JS_NODELIB_TARGET)
+    # Generate node.lib
+    execute_process(COMMAND ${CMAKE_AR} /def:${CMAKE_JS_NODELIB_DEF} /out:${CMAKE_JS_NODELIB_TARGET} ${CMAKE_STATIC_LINKER_FLAGS})
+endif()

examples/addon.node/README.md

@@ -14,14 +14,14 @@ npm install
 Make sure it is in the project root directory and compiled with make-js.
 
 ```shell
-npx cmake-js compile -T whisper-addon
+npx cmake-js compile -T whisper-addon -B Release
 ```
 
 For Electron addon and cmake-js options, you can see [cmake-js](https://github.com/cmake-js/cmake-js) and make very few configuration changes.
 
 > Such as appointing special cmake path:
 > ```shell
-> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon
+> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -B Release
 > ```
 
 ## Run

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

@@ -0,0 +1,23 @@
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
+
+const whisperParamsMock = {
+  language: "en",
+  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
+};
+
+describe("Run whisper.node", () => {
+    test("it should receive a non-empty value", async () => {
+        let result = await whisperAsync(whisperParamsMock);
+
+        expect(result.length).toBeGreaterThan(0);
+    }, 10000);
+});
+

examples/addon.node/addon.cpp

@@ -1,14 +1,13 @@
+#include "napi.h"
+#include "common.h"
+
+#include "whisper.h"
+
 #include <string>
 #include <thread>
 #include <vector>
 #include <cmath>
-
-#include "napi.h"
-
-#define DR_WAV_IMPLEMENTATION
-#include "dr_wav.h"
-
-#include "whisper.h"
+#include <cstdint>
 
 struct whisper_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
@@ -43,7 +42,7 @@ struct whisper_params {
     std::string model    = "../../ggml-large.bin";
 
     std::vector<std::string> fname_inp = {};
-    std::vector<std::string> fname_outp = {};
+    std::vector<std::string> fname_out = {};
 };
 
 struct whisper_print_user_data {
@@ -73,7 +72,7 @@ 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)));
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+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;
 
@@ -142,7 +141,6 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
 }
 
 int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
-
     if (params.fname_inp.empty()) {
         fprintf(stderr, "error: no input files specified\n");
         return 2;
@@ -162,109 +160,16 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
-        const auto fname_outp = f < (int)params.fname_outp.size() && !params.fname_outp[f].empty() ? params.fname_outp[f] : 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];
 
         std::vector<float> pcmf32; // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // WAV input
-        {
-            drwav wav;
-            std::vector<uint8_t> wav_data; // used for pipe input from stdin
-
-            if (fname_inp == "-") {
-                {
-                    uint8_t buf[1024];
-                    while (true)
-                    {
-                        const size_t n = fread(buf, 1, sizeof(buf), stdin);
-                        if (n == 0) {
-                            break;
-                        }
-                        wav_data.insert(wav_data.end(), buf, buf + n);
-                    }
-                }
-
-                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-                    fprintf(stderr, "error: failed to open WAV file from stdin\n");
-                    return 4;
-                }
-
-                fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
-            }
-            else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
-                fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
-                return 5;
-            }
-
-           if (wav.channels != 1 && wav.channels != 2) {
-               fprintf(stderr, "error: WAV file '%s' must be mono or stereo\n", fname_inp.c_str());
-               return 6;
-           }
-
-           if (params.diarize && wav.channels != 2 && params.no_timestamps == false) {
-               fprintf(stderr, "error: WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", fname_inp.c_str());
-               return 6;
-           }
-
-           if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-               fprintf(stderr, "error: WAV file '%s' must be %i kHz\n", fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
-               return 8;
-           }
-
-           if (wav.bitsPerSample != 16) {
-               fprintf(stderr, "error: WAV file '%s' must be 16-bit\n", fname_inp.c_str());
-               return 9;
-           }
-
-            const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
-
-            std::vector<int16_t> pcm16;
-            pcm16.resize(n*wav.channels);
-            drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
-            drwav_uninit(&wav);
-
-            // convert to mono, float
-            pcmf32.resize(n);
-            if (wav.channels == 1) {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[i])/32768.0f;
-                }
-            } else {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-                }
-            }
-
-            if (params.diarize) {
-                // convert to stereo, float
-                pcmf32s.resize(2);
-
-                pcmf32s[0].resize(n);
-                pcmf32s[1].resize(n);
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
-                    pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
-                }
-            }
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
         }
 
         // print system information
@@ -322,8 +227,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            wparams.initial_prompt   = params.prompt.c_str();
 
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
@@ -339,7 +243,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -371,51 +275,64 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
     return 0;
 }
 
-Napi::Object whisper(const Napi::CallbackInfo& info) {
-    Napi::Env env = info.Env();
-    if (info.Length() <= 0 || !info[0].IsObject()) {
-        Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
-    }
-    whisper_params params;
-    std::vector<std::vector<std::string>> result;
+class Worker : public Napi::AsyncWorker {
+ public:
+  Worker(Napi::Function& callback, whisper_params params)
+      : Napi::AsyncWorker(callback), params(params) {}
 
-    Napi::Object whisper_params = info[0].As<Napi::Object>();
-    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>();
-
-    params.language = language;
-    params.model = model;
-    params.fname_inp.emplace_back(input);
-
-    // run model
+  void Execute() override {
     run(params, result);
+  }
 
-    fprintf(stderr, "RESULT:\n");
-    for (auto sentence:result) {
-        fprintf(stderr, "t0: %s, t1: %s, content: %s \n",
-                sentence[0].c_str(), sentence[1].c_str(), sentence[2].c_str());
+  void OnOK() override {
+    Napi::HandleScope scope(Env());
+    Napi::Object res = Napi::Array::New(Env(), result.size());
+    for (uint64_t i = 0; i < result.size(); ++i) {
+      Napi::Object tmp = Napi::Array::New(Env(), 3);
+      for (uint64_t j = 0; j < 3; ++j) {
+        tmp[j] = Napi::String::New(Env(), result[i][j]);
+      }
+      res[i] = tmp;
     }
+    Callback().Call({Env().Null(), res});
+  }
 
-    Napi::Object res = Napi::Array::New(env, result.size());
-    for (u_int32_t i = 0; i < result.size(); ++i) {
-        Napi::Object tmp = Napi::Array::New(env, 3);
-        for (u_int32_t j = 0; j < 3; ++j) {
-            tmp[j] = Napi::String::New(env, result[i][j]);
-        }
-        res[i] = tmp;
-    }
+ private:
+  whisper_params params;
+  std::vector<std::vector<std::string>> result;
+};
 
-    return res;
+
+
+Napi::Value whisper(const Napi::CallbackInfo& info) {
+  Napi::Env env = info.Env();
+  if (info.Length() <= 0 || !info[0].IsObject()) {
+    Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
+  }
+  whisper_params params;
+
+  Napi::Object whisper_params = info[0].As<Napi::Object>();
+  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>();
+
+  params.language = language;
+  params.model = model;
+  params.fname_inp.emplace_back(input);
+
+  Napi::Function callback = info[1].As<Napi::Function>();
+  Worker* worker = new Worker(callback, params);
+  worker->Queue();
+  return env.Undefined();
 }
 
 
 Napi::Object Init(Napi::Env env, Napi::Object exports) {
-    exports.Set(
-            Napi::String::New(env, "whisper"),
-            Napi::Function::New(env, whisper)
-    );
-    return exports;
+  exports.Set(
+      Napi::String::New(env, "whisper"),
+      Napi::Function::New(env, whisper)
+  );
+  return exports;
 }
 
 NODE_API_MODULE(whisper, Init);

examples/addon.node/index.js

@@ -1,27 +1,36 @@
-const path = require('path');
-const { whisper } = require(path.join(__dirname, '../../build/Release/whisper-addon'));
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
 
 const whisperParams = {
-    language: 'en',
-    model: path.join(__dirname, '../../models/ggml-base.en.bin'),
-    fname_inp: '',
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: "../../samples/jfk.wav",
 };
 
 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
-    arguments.reduce((pre, item) => {
-        if (item.startsWith("--")) {
-            return [...pre, item.slice(2).split("=")];
-        }
-        return pre;
-    }, []),
+  arguments.reduce((pre, item) => {
+    if (item.startsWith("--")) {
+      return [...pre, item.slice(2).split("=")];
+    }
+    return pre;
+  }, [])
 );
 
 for (const key in params) {
-    if (whisperParams.hasOwnProperty(key)) {
-        whisperParams[key] = params[key];
-    }
+  if (whisperParams.hasOwnProperty(key)) {
+    whisperParams[key] = params[key];
+  }
 }
 
-console.log('whisperParams =', whisperParams);
-console.log(whisper(whisperParams));
+console.log("whisperParams =", whisperParams);
+
+whisperAsync(whisperParams).then((result) => {
+  console.log(`Result from whisper: ${result}`);
+});

examples/addon.node/package.json

@@ -5,8 +5,12 @@
   "main": "index.js",
   "author": "Qanhe Chen",
   "license": "MIT",
+  "scripts": {
+    "test": "jest"
+  },
   "devDependencies": {
     "cmake-js": "^7.1.1",
+    "jest": "^29.4.0",
     "node-addon-api": "^5.0.0"
   }
 }

examples/bench.wasm/CMakeLists.txt

@@ -31,9 +31,9 @@ endif()
 set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
     -s USE_PTHREADS=1 \
-    -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1024MB \
-    -s TOTAL_MEMORY=1024MB \
+    -s PTHREAD_POOL_SIZE_STRICT=0 \
+    -s INITIAL_MEMORY=2000MB \
+    -s TOTAL_MEMORY=2000MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/bench.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use and click the "Bench" button.<br>
@@ -44,11 +53,18 @@
 
             <div id="model-whisper">
                 Whisper model: <span id="model-whisper-status"></span>
-                <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
-                <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
-                <span id="fetch-whisper-progress"></span>
-
+                <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
+                <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
                 <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
+                <button id="fetch-whisper-small-en-q5_1"  onclick="loadWhisper('small-en-q5_1')">small.en (Q5_1, 182 MB)</button>
+                <button id="fetch-whisper-medium-en-q5_0" onclick="loadWhisper('medium-en-q5_0')">medium.en (Q5_0, 515 MB)</button>
+                <button id="fetch-whisper-large-q5_0"     onclick="loadWhisper('large-q5_0')">large (Q5_0, 1030 MB)</button>
+                <span id="fetch-whisper-progress"></span>
             </div>
 
             <br>
@@ -160,6 +176,14 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
                 document.getElementById('whisper-file'         ).style.display = 'none';
                 document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
             }
@@ -168,19 +192,42 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
+                    'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
+                    'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
+                    'large-q5_0':    'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+                    'small.en': 466,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
+                    'small-en-q5_1':  182,
+                    'medium-en-q5_0': 515,
+                    'large-q5_0':     1030,
                 };
 
                 let url     = urls[model];
                 let dst     = 'whisper.bin';
                 let size_mb = sizes[model];
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny-en').style.display  = 'none';
+                document.getElementById('fetch-whisper-base-en').style.display  = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
+                document.getElementById('whisper-file'        ).style.display = 'none';
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -190,9 +237,18 @@
 
                 cbCancel = function() {
                     var el;
-                    el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
+                    el = document.getElementById('fetch-whisper-tiny-en');  if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en');  if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en-q5_1' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-medium-en-q5_0'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-large-q5_0'    ); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('whisper-file'        ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
                 };
 
                 loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);

examples/bench/bench.cpp

@@ -44,13 +44,13 @@ 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, "                           %-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());
@@ -69,12 +69,49 @@ int whisper_bench_encoder(const whisper_params & params) {
         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);
         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 encode model: %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 encode model: %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 model: %d\n", ret);
+        return 4;
+    }
+
+    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 encode model: %d\n", ret);
+            return 4;
+        }
+    }
+
+    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 encode model: %d\n", ret);
+            return 4;
+        }
+    }
+
     whisper_print_timings(ctx);
     whisper_free(ctx);
 
@@ -103,7 +140,7 @@ 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/CMakeLists.txt

@@ -11,6 +11,7 @@ add_executable(${TARGET}
 include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE
+    common
     whisper
     )
 

examples/command.wasm/emscripten.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "common.h"
 #include "whisper.h"
 
 #include <emscripten.h>
@@ -27,92 +28,11 @@ std::string g_transcribed   = "";
 
 std::vector<float> g_pcmf32;
 
-static std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-static void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-// compute similarity between two strings using Levenshtein distance
-static float similarity(const std::string & s0, const std::string & s1) {
-    const size_t len0 = s0.size() + 1;
-    const size_t len1 = s1.size() + 1;
-
-    std::vector<int> col(len1, 0);
-    std::vector<int> prevCol(len1, 0);
-
-    for (size_t i = 0; i < len1; i++) {
-        prevCol[i] = i;
-    }
-
-    for (size_t i = 0; i < len0; i++) {
-        col[0] = i;
-        for (size_t j = 1; j < len1; j++) {
-            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (s0[i - 1] == s1[j - 1] ? 0 : 1));
-        }
-        col.swap(prevCol);
-    }
-
-    const float dist = prevCol[len1 - 1];
-
-    return 1.0f - (dist / std::max(s0.size(), s1.size()));
-}
-
 void command_set_status(const std::string & status) {
     std::lock_guard<std::mutex> lock(g_mutex);
     g_status = status;
 }
 
-bool command_vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (size_t i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string command_transcribe(whisper_context * ctx, const whisper_full_params & wparams, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -155,7 +75,7 @@ void command_get_audio(int ms, int sample_rate, std::vector<float> & audio) {
     const int64_t n_samples = (ms * sample_rate) / 1000;
 
     int64_t n_take = 0;
-    if (g_pcmf32.size() < n_samples) {
+    if (n_samples > (int) g_pcmf32.size()) {
         n_take = g_pcmf32.size();
     } else {
         n_take = n_samples;
@@ -187,7 +107,6 @@ void command_main(size_t index) {
 
     printf("command: using %d threads\n", wparams.n_threads);
 
-    bool is_running   = true;
     bool have_prompt  = false;
     bool ask_prompt   = true;
     bool print_energy = false;
@@ -233,7 +152,7 @@ void command_main(size_t index) {
         {
             command_get_audio(vad_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
 
-            if (command_vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
                 command_set_status("Speech detected! Processing ...");
 

examples/command.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use, click the "Start" button and follow the instructions.
@@ -45,6 +54,10 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <!--
@@ -162,11 +175,17 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
                 };
 
                 let url     = urls[model];
@@ -177,6 +196,10 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
+
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -188,6 +211,10 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
+
                     el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
                 };
 

examples/command/CMakeLists.txt

@@ -1,10 +1,9 @@
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # command
     set(TARGET command)
     add_executable(${TARGET} command.cpp)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/command/command.cpp

@@ -6,11 +6,10 @@
 // ref: https://github.com/ggerganov/whisper.cpp/issues/171
 //
 
+#include "common-sdl.h"
+#include "common.h"
 #include "whisper.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
 #include <sstream>
 #include <cassert>
 #include <cstdio>
@@ -110,309 +109,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    bool       m_running = false;
-    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;
-};
-
-audio_async::audio_async(int len_ms) {
-    m_len_ms = len_ms;
-}
-
-audio_async::~audio_async() {
-    if (m_dev_id_in) {
-        SDL_CloseAudioDevice(m_dev_id_in);
-    }
-}
-
-bool audio_async::init(int capture_id, int sample_rate) {
-    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
-
-    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
-        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
-        return false;
-    }
-
-    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
-
-    {
-        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
-        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
-        for (int i = 0; i < nDevices; i++) {
-            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
-        }
-    }
-
-    SDL_AudioSpec capture_spec_requested;
-    SDL_AudioSpec capture_spec_obtained;
-
-    SDL_zero(capture_spec_requested);
-    SDL_zero(capture_spec_obtained);
-
-    capture_spec_requested.freq     = sample_rate;
-    capture_spec_requested.format   = AUDIO_F32;
-    capture_spec_requested.channels = 1;
-    capture_spec_requested.samples  = 1024;
-    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
-        audio_async * audio = (audio_async *) userdata;
-        audio->callback(stream, len);
-    };
-    capture_spec_requested.userdata = this;
-
-    if (capture_id >= 0) {
-        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    } else {
-        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
-        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    }
-
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-        m_dev_id_in = 0;
-
-        return false;
-    } else {
-        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
-        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
-        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
-                capture_spec_requested.format);
-        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
-                capture_spec_requested.channels);
-        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
-    }
-
-    m_sample_rate = capture_spec_obtained.freq;
-
-    m_audio.resize((m_sample_rate*m_len_ms)/1000);
-
-    return true;
-}
-
-bool audio_async::resume() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
-        return false;
-    }
-
-    if (m_running) {
-        fprintf(stderr, "%s: already running!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 0);
-
-    m_running = true;
-
-    return true;
-}
-
-bool audio_async::pause() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: already paused!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 1);
-
-    m_running = false;
-
-    return true;
-}
-
-bool audio_async::clear() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return false;
-    }
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        m_audio_pos = 0;
-        m_audio_len = 0;
-    }
-
-    return true;
-}
-
-// callback to be called by SDL
-void audio_async::callback(uint8_t * stream, int len) {
-    if (!m_running) {
-        return;
-    }
-
-    const size_t n_samples = len / sizeof(float);
-
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
-
-    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (m_audio_pos + n_samples > m_audio.size()) {
-            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));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
-        } else {
-            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
-        }
-    }
-}
-
-void audio_async::get(int ms, std::vector<float> & result) {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
-        return;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return;
-    }
-
-    result.clear();
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (ms <= 0) {
-            ms = m_len_ms;
-        }
-
-        size_t n_samples = (m_sample_rate * ms) / 1000;
-        if (n_samples > m_audio_len) {
-            n_samples = m_audio_len;
-        }
-
-        result.resize(n_samples);
-
-        int s0 = m_audio_pos - n_samples;
-        if (s0 < 0) {
-            s0 += m_audio.size();
-        }
-
-        if (s0 + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - s0;
-
-            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
-            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
-        } else {
-            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
-        }
-    }
-}
-
-///////////////////////////
-
-std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (int i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -467,31 +163,6 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     return result;
 }
 
-// compute similarity between two strings using Levenshtein distance
-float similarity(const std::string & s0, const std::string & s1) {
-    const size_t len0 = s0.size() + 1;
-    const size_t len1 = s1.size() + 1;
-
-    std::vector<int> col(len1, 0);
-    std::vector<int> prevCol(len1, 0);
-
-    for (size_t i = 0; i < len1; i++) {
-        prevCol[i] = i;
-    }
-
-    for (size_t i = 0; i < len0; i++) {
-        col[0] = i;
-        for (size_t j = 1; j < len1; j++) {
-            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (s0[i - 1] == s1[j - 1] ? 0 : 1));
-        }
-        col.swap(prevCol);
-    }
-
-    const float dist = prevCol[len1 - 1];
-
-    return 1.0f - (dist / std::max(s0.size(), s1.size()));
-}
-
 std::vector<std::string> read_allowed_commands(const std::string & fname) {
     std::vector<std::string> allowed_commands;
 
@@ -502,7 +173,7 @@ std::vector<std::string> read_allowed_commands(const std::string & fname) {
 
     std::string line;
     while (std::getline(ifs, line)) {
-        line = trim(line);
+        line = ::trim(line);
         if (line.empty()) {
             continue;
         }
@@ -526,23 +197,6 @@ std::vector<std::string> get_words(const std::string &txt) {
     return words;
 }
 
-// returns true if no exit event was received
-bool process_sdl_events() {
-    SDL_Event event;
-    while (SDL_PollEvent(&event)) {
-        switch (event.type) {
-            case SDL_QUIT:
-                {
-                    return false;
-                } break;
-            default:
-                break;
-        }
-    }
-
-    return true;
-}
-
 // command-list mode
 // guide the transcription to match the most likely command from a provided list
 int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
@@ -634,14 +288,14 @@ int process_command_list(struct whisper_context * ctx, audio_async &audio, const
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
 
         audio.get(2000, pcmf32_cur);
 
-        if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+        if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
             fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
             const auto t_start = std::chrono::high_resolution_clock::now();
@@ -775,7 +429,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
@@ -791,7 +445,7 @@ int always_prompt_transcription(struct whisper_context * ctx, audio_async & audi
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 int64_t t_ms = 0;
@@ -854,7 +508,7 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        is_running = process_sdl_events();
+        is_running = sdl_poll_events();
 
         // delay
         std::this_thread::sleep_for(std::chrono::milliseconds(100));
@@ -870,7 +524,7 @@ int process_general_transcription(struct whisper_context * ctx, audio_async &aud
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 int64_t t_ms = 0;

examples/common-ggml.cpp

@@ -0,0 +1,246 @@
+#include "common-ggml.h"
+
+#include <regex>
+#include <map>
+
+static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
+    {"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
+    {"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
+    {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
+    {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
+    {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+};
+
+void ggml_print_ftypes(FILE * fp) {
+    for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
+        fprintf(fp, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+    }
+}
+
+enum ggml_ftype ggml_parse_ftype(const char * str) {
+    enum ggml_ftype ftype;
+    if (str[0] == 'q') {
+        const auto it = GGML_FTYPE_MAP.find(str);
+        if (it == GGML_FTYPE_MAP.end()) {
+            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
+            return GGML_FTYPE_UNKNOWN;
+        }
+        ftype = it->second;
+    } else {
+        ftype = (enum ggml_ftype) atoi(str);
+    }
+
+    return ftype;
+}
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip) {
+
+    ggml_type qtype = GGML_TYPE_F32;
+
+    switch (ftype) {
+        case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
+        case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
+        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_UNKNOWN:
+        case GGML_FTYPE_ALL_F32:
+        case GGML_FTYPE_MOSTLY_F16:
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_Q2_K:
+        case GGML_FTYPE_MOSTLY_Q3_K:
+        case GGML_FTYPE_MOSTLY_Q4_K:
+        case GGML_FTYPE_MOSTLY_Q5_K:
+        case GGML_FTYPE_MOSTLY_Q6_K:
+                {
+                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
+                    return false;
+                }
+    };
+
+    if (!ggml_is_quantized(qtype)) {
+        fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
+        return false;
+    }
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector<float> work;
+
+    std::vector<uint8_t>     data_u8;
+    std::vector<ggml_fp16_t> data_f16;
+    std::vector<float>       data_f32;
+
+    std::vector<int64_t> hist_all(1 << 4, 0);
+
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ttype;
+
+        finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+        finp.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+        if (finp.eof()) {
+            break;
+        }
+
+        int32_t nelements = 1;
+        int32_t ne[4] = { 1, 1, 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            nelements *= ne[i];
+        }
+
+        std::string name(length, 0);
+        finp.read (&name[0], length);
+
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
+
+        bool quantize = false;
+
+        // check if we should quantize this tensor
+        for (const auto & s : to_quant) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // check if we should skip this tensor
+        for (const auto & s : to_skip) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = false;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (n_dims == 2);
+
+        if (quantize) {
+            if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
+                fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                return false;
+            }
+
+            if (ttype == GGML_TYPE_F16) {
+                data_f16.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                data_f32.resize(nelements);
+                for (int i = 0; i < nelements; ++i) {
+                    data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                }
+            } else {
+                data_f32.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+            }
+
+            ttype = qtype;
+        } else {
+            const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+            data_u8.resize(nelements*bpe);
+            finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+        }
+
+        fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+        fout.write(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+        for (int i = 0; i < n_dims; ++i) {
+            fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+        fout.write(&name[0], length);
+
+        if (quantize) {
+            work.resize(nelements); // for quantization
+
+            size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 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:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_I8:
+                case GGML_TYPE_I16:
+                case GGML_TYPE_I32:
+                case GGML_TYPE_Q8_1:
+                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:
+                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));
+                        return false;
+                    }
+            }
+
+            fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+            total_size_new += cur_size;
+
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
+        } else {
+            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+            fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+            total_size_new += data_u8.size();
+        }
+
+        total_size_org += nelements * sizeof(float);
+    }
+
+    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
+
+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
+    return true;
+}

examples/common-ggml.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <fstream>
+#include <vector>
+#include <string>
+
+enum ggml_ftype ggml_parse_ftype(const char * str);
+
+void ggml_print_ftypes(FILE * fp = stderr);
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip);

examples/common-sdl.cpp

@@ -0,0 +1,226 @@
+#include "common-sdl.h"
+
+audio_async::audio_async(int len_ms) {
+    m_len_ms = len_ms;
+
+    m_running = false;
+}
+
+audio_async::~audio_async() {
+    if (m_dev_id_in) {
+        SDL_CloseAudioDevice(m_dev_id_in);
+    }
+}
+
+bool audio_async::init(int capture_id, int sample_rate) {
+    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
+
+    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
+        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
+        return false;
+    }
+
+    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
+
+    {
+        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
+        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
+        for (int i = 0; i < nDevices; i++) {
+            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
+        }
+    }
+
+    SDL_AudioSpec capture_spec_requested;
+    SDL_AudioSpec capture_spec_obtained;
+
+    SDL_zero(capture_spec_requested);
+    SDL_zero(capture_spec_obtained);
+
+    capture_spec_requested.freq     = sample_rate;
+    capture_spec_requested.format   = AUDIO_F32;
+    capture_spec_requested.channels = 1;
+    capture_spec_requested.samples  = 1024;
+    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
+        audio_async * audio = (audio_async *) userdata;
+        audio->callback(stream, len);
+    };
+    capture_spec_requested.userdata = this;
+
+    if (capture_id >= 0) {
+        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
+        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
+    } else {
+        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
+        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
+    }
+
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
+        m_dev_id_in = 0;
+
+        return false;
+    } else {
+        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
+        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
+        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
+                capture_spec_requested.format);
+        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
+                capture_spec_requested.channels);
+        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
+    }
+
+    m_sample_rate = capture_spec_obtained.freq;
+
+    m_audio.resize((m_sample_rate*m_len_ms)/1000);
+
+    return true;
+}
+
+bool audio_async::resume() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
+        return false;
+    }
+
+    if (m_running) {
+        fprintf(stderr, "%s: already running!\n", __func__);
+        return false;
+    }
+
+    SDL_PauseAudioDevice(m_dev_id_in, 0);
+
+    m_running = true;
+
+    return true;
+}
+
+bool audio_async::pause() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
+        return false;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: already paused!\n", __func__);
+        return false;
+    }
+
+    SDL_PauseAudioDevice(m_dev_id_in, 1);
+
+    m_running = false;
+
+    return true;
+}
+
+bool audio_async::clear() {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
+        return false;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: not running!\n", __func__);
+        return false;
+    }
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        m_audio_pos = 0;
+        m_audio_len = 0;
+    }
+
+    return true;
+}
+
+// callback to be called by SDL
+void audio_async::callback(uint8_t * stream, int len) {
+    if (!m_running) {
+        return;
+    }
+
+    const size_t n_samples = len / sizeof(float);
+
+    m_audio_new.resize(n_samples);
+    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
+
+    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        if (m_audio_pos + n_samples > m_audio.size()) {
+            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));
+
+            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
+            m_audio_len = m_audio.size();
+        } else {
+            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
+
+            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
+            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
+        }
+    }
+}
+
+void audio_async::get(int ms, std::vector<float> & result) {
+    if (!m_dev_id_in) {
+        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
+        return;
+    }
+
+    if (!m_running) {
+        fprintf(stderr, "%s: not running!\n", __func__);
+        return;
+    }
+
+    result.clear();
+
+    {
+        std::lock_guard<std::mutex> lock(m_mutex);
+
+        if (ms <= 0) {
+            ms = m_len_ms;
+        }
+
+        size_t n_samples = (m_sample_rate * ms) / 1000;
+        if (n_samples > m_audio_len) {
+            n_samples = m_audio_len;
+        }
+
+        result.resize(n_samples);
+
+        int s0 = m_audio_pos - n_samples;
+        if (s0 < 0) {
+            s0 += m_audio.size();
+        }
+
+        if (s0 + n_samples > m_audio.size()) {
+            const size_t n0 = m_audio.size() - s0;
+
+            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
+            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
+        } else {
+            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
+        }
+    }
+}
+
+bool sdl_poll_events() {
+    SDL_Event event;
+    while (SDL_PollEvent(&event)) {
+        switch (event.type) {
+            case SDL_QUIT:
+                {
+                    return false;
+                } break;
+            default:
+                break;
+        }
+    }
+
+    return true;
+}

examples/common-sdl.h

@@ -0,0 +1,50 @@
+#pragma once
+
+#include <SDL.h>
+#include <SDL_audio.h>
+
+#include <atomic>
+#include <cstdint>
+#include <vector>
+#include <mutex>
+
+//
+// SDL Audio capture
+//
+
+class audio_async {
+public:
+    audio_async(int len_ms);
+    ~audio_async();
+
+    bool init(int capture_id, int sample_rate);
+
+    // start capturing audio via the provided SDL callback
+    // keep last len_ms seconds of audio in a circular buffer
+    bool resume();
+    bool pause();
+    bool clear();
+
+    // callback to be called by SDL
+    void callback(uint8_t * stream, int len);
+
+    // get audio data from the circular buffer
+    void get(int ms, std::vector<float> & audio);
+
+private:
+    SDL_AudioDeviceID m_dev_id_in = 0;
+
+    int m_len_ms = 0;
+    int m_sample_rate = 0;
+
+    std::atomic_bool m_running;
+    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;
+};
+
+// Return false if need to quit
+bool sdl_poll_events();

examples/common.cpp

@@ -0,0 +1,809 @@
+#define _USE_MATH_DEFINES // for M_PI
+
+#include "common.h"
+
+// third-party utilities
+// use your favorite implementations
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <regex>
+#include <locale>
+#include <codecvt>
+#include <sstream>
+
+#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(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = 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 == "-p" || arg == "--prompt") {
+            params.prompt = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-n" || arg == "--n_predict") {
+            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_k") {
+            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_p") {
+            params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--temp") {
+            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(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-m" || arg == "--model") {
+            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") {
+            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]);
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        } 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);
+        }
+    }
+
+    return true;
+}
+
+void gpt_print_usage(int /*argc*/, char ** argv, const gpt_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, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    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, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string gpt_random_prompt(std::mt19937 & rng) {
+    const int r = rng() % 10;
+    switch (r) {
+        case 0: return "So";
+        case 1: return "Once upon a time";
+        case 2: return "When";
+        case 3: return "The";
+        case 4: return "After";
+        case 5: return "If";
+        case 6: return "import";
+        case 7: return "He";
+        case 8: return "She";
+        case 9: return "They";
+        default: return "To";
+    }
+
+    return "The";
+}
+
+std::string trim(const std::string & s) {
+    std::regex e("^\\s+|\\s+$");
+    return std::regex_replace(s, e, "");
+}
+
+std::string replace(const std::string & s, const std::string & from, const std::string & to) {
+    std::string result = s;
+    size_t pos = 0;
+    while ((pos = result.find(from, pos)) != std::string::npos) {
+        result.replace(pos, from.length(), to);
+        pos += to.length();
+    }
+    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;
+
+    // read file into string
+    std::string json;
+    {
+        std::ifstream ifs(fname);
+        if (!ifs) {
+            fprintf(stderr, "Failed to open %s\n", fname.c_str());
+            exit(1);
+        }
+
+        json = std::string((std::istreambuf_iterator<char>(ifs)),
+                (std::istreambuf_iterator<char>()));
+    }
+
+    if (json[0] != '{') {
+        return result;
+    }
+
+    // parse json
+    {
+        bool has_key  = false;
+        bool in_token = false;
+
+        std::string str_key = "";
+        std::string str_val = "";
+
+        int n = json.size();
+        for (int i = 1; i < n; ++i) {
+            if (!in_token) {
+                if (json[i] == ' ') continue;
+                if (json[i] == '"') {
+                    in_token = true;
+                    continue;
+                }
+            } else {
+                if (json[i] == '\\' && i+1 < n) {
+                    if (has_key == false) {
+                        str_key += json[i];
+                    } else {
+                        str_val += json[i];
+                    }
+                    ++i;
+                } else if (json[i] == '"') {
+                    if (has_key == false) {
+                        has_key = true;
+                        ++i;
+                        while (json[i] == ' ') ++i;
+                        ++i; // :
+                        while (json[i] == ' ') ++i;
+                        if (json[i] != '\"') {
+                            while (json[i] != ',' && json[i] != '}') {
+                                str_val += json[i++];
+                            }
+                            has_key = false;
+                        } else {
+                            in_token = true;
+                            continue;
+                        }
+                    } else {
+                        has_key = false;
+                    }
+
+                    str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
+                    str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
+                    str_key = ::replace(str_key, "\\\"",    "\""); // \\\"   -> "
+
+                    try {
+                        result[str_key] = std::stoi(str_val);
+                    } catch (...) {
+                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
+
+                    }
+                    str_key = "";
+                    str_val = "";
+                    in_token = false;
+                    continue;
+                }
+                if (has_key == false) {
+                    str_key += json[i];
+                } else {
+                    str_val += json[i];
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
+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) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+
+        // 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 += std::regex_replace(token, escape, R"(\$&)");
+            }
+
+            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();
+            }
+            // Remaining text without special tokens will be handled below.
+        }
+
+        gpt_split_words(str, words);
+    }
+
+    // find the longest token that forms each word in words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        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 + 1;
+                    break;
+                }
+                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++;
+                }
+            }
+        }
+    }
+
+    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());
+
+    vocab.token_to_id = ::json_parse(fname);
+
+    for (const auto & kv : vocab.token_to_id) {
+        vocab.id_to_token[kv.second] = kv.first;
+    }
+
+    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
+
+    // print the vocabulary
+    //for (auto kv : vocab.token_to_id) {
+    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
+    //}
+
+    return true;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[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++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    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
+
+    if (fname == "-") {
+        {
+            uint8_t buf[1024];
+            while (true)
+            {
+                const size_t n = fread(buf, 1, sizeof(buf), stdin);
+                if (n == 0) {
+                    break;
+                }
+                wav_data.insert(wav_data.end(), buf, buf + n);
+            }
+        }
+
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from stdin\n");
+            return false;
+        }
+
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    }
+    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
+        return false;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (stereo && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
+        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        return false;
+    }
+
+    if (wav.bitsPerSample != 16) {
+        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        return false;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[i])/32768.0f;
+        }
+    } else {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+        }
+    }
+
+    if (stereo) {
+        // convert to stereo, float
+        pcmf32s.resize(2);
+
+        pcmf32s[0].resize(n);
+        pcmf32s[1].resize(n);
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+            pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+        }
+    }
+
+    return true;
+}
+
+void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
+    const float rc = 1.0f / (2.0f * M_PI * cutoff);
+    const float dt = 1.0f / sample_rate;
+    const float alpha = dt / (rc + dt);
+
+    float y = data[0];
+
+    for (size_t i = 1; i < data.size(); i++) {
+        y = alpha * (y + data[i] - data[i - 1]);
+        data[i] = y;
+    }
+}
+
+bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
+    const int n_samples      = pcmf32.size();
+    const int n_samples_last = (sample_rate * last_ms) / 1000;
+
+    if (n_samples_last >= n_samples) {
+        // not enough samples - assume no speech
+        return false;
+    }
+
+    if (freq_thold > 0.0f) {
+        high_pass_filter(pcmf32, freq_thold, sample_rate);
+    }
+
+    float energy_all  = 0.0f;
+    float energy_last = 0.0f;
+
+    for (int i = 0; i < n_samples; i++) {
+        energy_all += fabsf(pcmf32[i]);
+
+        if (i >= n_samples - n_samples_last) {
+            energy_last += fabsf(pcmf32[i]);
+        }
+    }
+
+    energy_all  /= n_samples;
+    energy_last /= n_samples_last;
+
+    if (verbose) {
+        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
+    }
+
+    if (energy_last > vad_thold*energy_all) {
+        return false;
+    }
+
+    return true;
+}
+
+float similarity(const std::string & s0, const std::string & s1) {
+    const size_t len0 = s0.size() + 1;
+    const size_t len1 = s1.size() + 1;
+
+    std::vector<int> col(len1, 0);
+    std::vector<int> prevCol(len1, 0);
+
+    for (size_t i = 0; i < len1; i++) {
+        prevCol[i] = i;
+    }
+
+    for (size_t i = 0; i < len0; i++) {
+        col[0] = i;
+        for (size_t j = 1; j < len1; j++) {
+            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (i > 0 && s0[i - 1] == s1[j - 1] ? 0 : 1));
+        }
+        col.swap(prevCol);
+    }
+
+    const float dist = prevCol[len1 - 1];
+
+    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

@@ -0,0 +1,176 @@
+// Various helper functions and utilities
+
+#pragma once
+
+#include <string>
+#include <map>
+#include <vector>
+#include <random>
+#include <thread>
+
+#define COMMON_SAMPLE_RATE 16000
+
+//
+// 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 n_batch   = 8;   // batch size for prompt processing
+
+    // sampling parameters
+    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;
+
+    std::string model      = "models/gpt-2-117M/ggml-model.bin"; // model path
+    std::string prompt     = "";
+    std::string token_test = "";
+
+    bool    interactive      = false;
+    int32_t interactive_port = -1;
+
+    int32_t n_gpu_layers     = 0;
+};
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+//
+// Vocab utils
+//
+
+std::string trim(const std::string & s);
+
+std::string replace(
+        const std::string & s,
+        const std::string & from,
+        const std::string & to);
+
+struct gpt_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+    std::vector<std::string> special_tokens;
+
+    void add_special_token(const std::string & token);
+};
+
+// 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
+//
+// Regex (Python):
+// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
+//
+// Regex (C++):
+// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
+//
+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);
+
+// sample next token given probabilities for each embedding
+//
+//   - consider only the top K tokens
+//   - from them, consider only the top tokens with cumulative probability > P
+//
+// TODO: not sure if this implementation is correct
+// TODO: temperature is not implemented
+//
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double 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
+//
+
+// Read WAV audio file and store the PCM data into pcmf32
+// The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
+// If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
+bool read_wav(
+        const std::string & fname,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
+// Apply a high-pass frequency filter to PCM audio
+// Suppresses frequencies below cutoff Hz
+void high_pass_filter(
+        std::vector<float> & data,
+        float cutoff,
+        float sample_rate);
+
+// Basic voice activity detection (VAD) using audio energy adaptive threshold
+bool vad_simple(
+        std::vector<float> & pcmf32,
+        int   sample_rate,
+        int   last_ms,
+        float vad_thold,
+        float freq_thold,
+        bool  verbose);
+
+// 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/helpers.js

@@ -145,7 +145,15 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
                             var db = event.target.result;
                             var tx = db.transaction(['models'], 'readwrite');
                             var os = tx.objectStore('models');
-                            var rq = os.put(data, url);
+
+                            var rq = null;
+                            try {
+                                var rq = os.put(data, url);
+                            } catch (e) {
+                                cbPrint('loadRemote: failed to store "' + url + '" in the IndexedDB: \n' + e);
+                                cbCancel();
+                                return;
+                            }
 
                             rq.onsuccess = function (event) {
                                 cbPrint('loadRemote: "' + url + '" stored in the IndexedDB');
@@ -180,7 +188,6 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
 
     rq.onabort = function (event) {
         cbPrint('loadRemote: failed to open IndexedDB: abort');
-
+        cbCancel();
     };
 }
-

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 +1,458 @@
+#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;
+
+    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 == "-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, "  -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 * ctx = whisper_init_from_file(params.model.c_str());
+    // 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/CMakeLists.txt

@@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)
 
 include(DefaultTargetOptions)
 
-target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/main/README.md

@@ -31,6 +31,7 @@ options:
   -osrt,     --output-srt        [false  ] output result in a srt file
   -owts,     --output-words      [false  ] output script for generating 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

examples/main/main.cpp

@@ -1,9 +1,6 @@
-#include "whisper.h"
+#include "common.h"
 
-// third-party utilities
-// use your favorite implementations
-#define DR_WAV_IMPLEMENTATION
-#include "dr_wav.h"
+#include "whisper.h"
 
 #include <cmath>
 #include <fstream>
@@ -11,6 +8,11 @@
 #include <string>
 #include <thread>
 #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.
@@ -57,35 +59,49 @@ 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      =  5;
+    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;
 
-    bool speed_up       = false;
-    bool translate      = false;
-    bool diarize        = 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 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_lrc      = false;
+    bool print_special   = false;
+    bool print_colors    = false;
+    bool print_progress  = false;
+    bool no_timestamps   = false;
+    bool log_score       = false;
 
-    std::string language = "en";
+    std::string language  = "en";
     std::string prompt;
-    std::string model    = "models/ggml-base.en.bin";
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    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_outp = {};
+    std::vector<std::string> fname_out = {};
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -94,6 +110,11 @@ 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 == "-"){
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
         if (arg[0] != '-') {
             params.fname_inp.push_back(arg);
             continue;
@@ -103,36 +124,45 @@ 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 == "-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 == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
-        else if (arg == "-of"   || arg == "--output-file")    { params.fname_outp.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 == "--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 == "-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 {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -156,29 +186,38 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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, "  -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");
     fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
+    fprintf(stderr, "  -olrc,     --output-lrc        [%-7s] output result in a lrc file\n",                    params.output_lrc ? "true" : "false");
     fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    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, "  -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");
     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 ? "false" : "true");
+    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, "  -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, "\n");
 }
 
@@ -186,9 +225,51 @@ struct whisper_print_user_data {
     const whisper_params * params;
 
     const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
 };
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+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;
 
@@ -196,8 +277,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
 
     std::string speaker = "";
 
-    int64_t t0;
-    int64_t t1;
+    int64_t t0 = 0;
+    int64_t t1 = 0;
 
     // print the last n_new segments
     const int s0 = n_segments - n_new;
@@ -217,28 +298,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
         }
 
         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) {
@@ -263,6 +323,12 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
             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");
@@ -272,7 +338,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
     }
 }
 
-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);
@@ -284,13 +350,22 @@ 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);
@@ -306,15 +381,23 @@ 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);
@@ -328,16 +411,53 @@ 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;
 }
 
-bool output_csv(struct whisper_context * ctx, const char * fname) {
+char *escape_double_quotes_and_backslashes(const char *str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped[pos++] = '\\';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+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);
@@ -347,28 +467,199 @@ 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,";
+    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);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        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    << "\"\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_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) {
+    std::ofstream fout(fname);
+    int indent = 0;
+
+    auto doindent = [&]() {
+        for (int i = 0; i < indent; i++) fout << "\t";
+    };
+
+    auto start_arr = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": [\n";
+        indent++;
+    };
+
+    auto end_arr = [&](bool end) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "},\n");
+    };
+
+    auto start_obj = [&](const char *name) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end) {
+        indent--;
+        doindent();
+        fout << (end ? "}\n" : "},\n");
+    };
+
+    auto start_value = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": ";
+    };
+
+    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) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    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) {
+        start_value(name);
+        fout << (val ? "true" : "false");
+        end_value(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(nullptr);
+        value_s("systeminfo", whisper_print_system_info(), false);
+        start_obj("model");
+            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), 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(false);
+            start_obj("text");
+                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(false);
+            value_i("mels", whisper_model_n_mels(ctx), false);
+            value_i("ftype", whisper_model_ftype(ctx), true);
+        end_obj(false);
+        start_obj("params");
+            value_s("model", params.model.c_str(), false);
+            value_s("language", params.language.c_str(), false);
+            value_b("translate", params.translate, true);
+        end_obj(false);
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        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(nullptr);
+                    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(false);
+                    value_s("text", text, !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));
+            }
+
+        end_arr(true);
+    end_obj(true);
+    return true;
+}
+
 // 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);
 
-    // TODO: become parameter
-    static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
 
     fout << "#!/bin/bash" << "\n";
     fout << "\n";
@@ -394,6 +685,11 @@ 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];
@@ -402,13 +698,19 @@ 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) {
@@ -471,10 +773,51 @@ 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, 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);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    fout << "[by:whisper.cpp]\n";
+
+    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 t = whisper_full_get_segment_t0(ctx, i);
+
+        int64_t msec = t * 10;
+        int64_t min = msec / (1000 * 60);
+        msec = msec - min * (1000 * 60);
+        int64_t sec = msec / 1000;
+        msec = msec - sec * 1000;
+
+        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 = "";
+
+        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;
+}
+
 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;
     }
 
@@ -490,6 +833,12 @@ 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());
@@ -499,109 +848,19 @@ int main(int argc, char ** argv) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
+    // 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_outp = f < (int) params.fname_outp.size() && !params.fname_outp[f].empty() ? params.fname_outp[f] : 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];
 
-        std::vector<float> pcmf32; // mono-channel F32 PCM
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // WAV input
-        {
-            drwav wav;
-            std::vector<uint8_t> wav_data; // used for pipe input from stdin
-
-            if (fname_inp == "-") {
-                {
-                    uint8_t buf[1024];
-                    while (true)
-                    {
-                        const size_t n = fread(buf, 1, sizeof(buf), stdin);
-                        if (n == 0) {
-                            break;
-                        }
-                        wav_data.insert(wav_data.end(), buf, buf + n);
-                    }
-                }
-
-                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-                    fprintf(stderr, "error: failed to open WAV file from stdin\n");
-                    return 4;
-                }
-
-                fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
-            }
-            else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
-                fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
-                return 5;
-            }
-
-            if (wav.channels != 1 && wav.channels != 2) {
-                fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (params.diarize && wav.channels != 2 && params.no_timestamps == false) {
-                fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-                fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", argv[0], fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
-                return 8;
-            }
-
-            if (wav.bitsPerSample != 16) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", argv[0], fname_inp.c_str());
-                return 9;
-            }
-
-            const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
-
-            std::vector<int16_t> pcm16;
-            pcm16.resize(n*wav.channels);
-            drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
-            drwav_uninit(&wav);
-
-            // convert to mono, float
-            pcmf32.resize(n);
-            if (wav.channels == 1) {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[i])/32768.0f;
-                }
-            } else {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-                }
-            }
-
-            if (params.diarize) {
-                // convert to stereo, float
-                pcmf32s.resize(2);
-
-                pcmf32s[0].resize(n);
-                pcmf32s[1].resize(n);
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
-                    pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
-                }
-            }
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
         }
 
         // print system information
@@ -621,11 +880,15 @@ int main(int argc, char ** argv) {
                     fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                 }
             }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
+            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__, fname_inp.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");
@@ -643,6 +906,7 @@ int main(int argc, char ** argv) {
             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;
@@ -651,11 +915,14 @@ int main(int argc, char ** argv) {
             wparams.token_timestamps = params.output_wts || 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.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            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;
@@ -664,7 +931,7 @@ 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) {
@@ -672,13 +939,18 @@ int main(int argc, char ** argv) {
                 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;
+            }
+
             // example for abort mechanism
             // in this example, 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*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -697,34 +969,51 @@ int main(int argc, char ** argv) {
 
             // output to text file
             if (params.output_txt) {
-                const auto fname_txt = fname_outp + ".txt";
-                output_txt(ctx, fname_txt.c_str());
+                const auto fname_txt = fname_out + ".txt";
+                output_txt(ctx, fname_txt.c_str(), params, pcmf32s);
             }
 
             // output to VTT file
             if (params.output_vtt) {
-                const auto fname_vtt = fname_outp + ".vtt";
-                output_vtt(ctx, fname_vtt.c_str());
+                const auto fname_vtt = fname_out + ".vtt";
+                output_vtt(ctx, fname_vtt.c_str(), params, pcmf32s);
             }
 
             // output to SRT file
             if (params.output_srt) {
-                const auto fname_srt = fname_outp + ".srt";
-                output_srt(ctx, fname_srt.c_str(), params);
+                const auto fname_srt = fname_out + ".srt";
+                output_srt(ctx, fname_srt.c_str(), params, pcmf32s);
             }
 
             // output to WTS file
             if (params.output_wts) {
-                const auto fname_wts = fname_outp + ".wts";
-                output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);
+                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, pcmf32s);
             }
 
-	    // output to CSV file
+            // output to CSV file
             if (params.output_csv) {
-                const auto fname_csv = fname_outp + ".csv";
-                output_csv(ctx, fname_csv.c_str());
+                const auto fname_csv = fname_out + ".csv";
+                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, pcmf32s);
+            }
+
+            // output to LRC file
+            if (params.output_lrc) {
+                const auto fname_lrc = fname_out + ".lrc";
+                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/CMakeLists.txt

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

examples/quantize/README.md

@@ -0,0 +1,3 @@
+# quantize
+
+Tool for integer quantization of Whisper `ggml` model files

examples/quantize/quantize.cpp

@@ -0,0 +1,223 @@
+#include "ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (Whisper tiny)
+struct whisper_hparams {
+    int32_t n_vocab       = 51864;
+    int32_t n_audio_ctx   = 1500;
+    int32_t n_audio_state = 384;
+    int32_t n_audio_head  = 6;
+    int32_t n_audio_layer = 4;
+    int32_t n_text_ctx    = 448;
+    int32_t n_text_state  = 384;
+    int32_t n_text_head   = 6;
+    int32_t n_text_layer  = 4;
+    int32_t n_mels        = 80;
+    int32_t ftype         = 1;
+};
+
+struct whisper_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+// quantize a model
+bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    whisper_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        finp.read((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        finp.read((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        finp.read((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        finp.read((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        finp.read((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        finp.read((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        finp.read((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        finp.read((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        finp.read((char *) &hparams.ftype,         sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        fprintf(stderr, "%s: n_vocab       = %d\n", __func__, hparams.n_vocab);
+        fprintf(stderr, "%s: n_audio_ctx   = %d\n", __func__, hparams.n_audio_ctx);
+        fprintf(stderr, "%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
+        fprintf(stderr, "%s: n_audio_head  = %d\n", __func__, hparams.n_audio_head);
+        fprintf(stderr, "%s: n_audio_layer = %d\n", __func__, hparams.n_audio_layer);
+        fprintf(stderr, "%s: n_text_ctx    = %d\n", __func__, hparams.n_text_ctx);
+        fprintf(stderr, "%s: n_text_state  = %d\n", __func__, hparams.n_text_state);
+        fprintf(stderr, "%s: n_text_head   = %d\n", __func__, hparams.n_text_head);
+        fprintf(stderr, "%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
+        fprintf(stderr, "%s: n_mels        = %d\n", __func__, hparams.n_mels);
+        fprintf(stderr, "%s: ftype (src)   = %d\n", __func__, hparams.ftype);
+        fprintf(stderr, "%s: qntvr (src)   = %d\n", __func__, qntvr_src);
+        fprintf(stderr, "%s: ftype (dst)   = %d\n", __func__, ftype_dst);
+        fprintf(stderr, "%s: qntvr (dst)   = %d\n", __func__, GGML_QNT_VERSION);
+
+        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
+    {
+        whisper_filters filters;
+
+        finp.read ((char *) &filters.n_mel, sizeof(filters.n_mel));
+        fout.write((char *) &filters.n_mel, sizeof(filters.n_mel));
+        finp.read ((char *) &filters.n_fft, sizeof(filters.n_fft));
+        fout.write((char *) &filters.n_fft, sizeof(filters.n_fft));
+
+        filters.data.resize(filters.n_mel * filters.n_fft);
+        finp.read ((char *) filters.data.data(), filters.data.size() * sizeof(float));
+        fout.write((char *) filters.data.data(), filters.data.size() * sizeof(float));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        //if (n_vocab != hparams.n_vocab) {
+        //    fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+        //            __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+        //    return false;
+        //}
+
+        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[len] = '\0';
+
+            finp.read ((char *) word, len);
+            fout.write((char *) word, len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to not be quantized
+    const std::vector<std::string> to_skip = {
+        //"encoder.*",
+        "encoder.conv1.bias",
+        "encoder.conv2.bias",
+        "encoder.positional_embedding",
+        "decoder.positional_embedding",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, { ".*" }, to_skip)) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!whisper_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/stream.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use, click the "Start" button and start speaking
@@ -45,6 +54,10 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <!--
@@ -162,11 +175,17 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
                 };
 
                 let url     = urls[model];
@@ -177,6 +196,10 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
+
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -188,6 +211,10 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
+
                     el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
                 };