release : v1.3.0

Also, fallback only if more temperatures are available and if we are
at least 3 seconds before the end of the audio

.github/workflows/bindings-go.yml

@@ -0,0 +1,22 @@
+name: Bindings Tests (Go)
+on:
+  push:
+    paths:
+      - bindings/go/**
+      - whisper.h
+  pull_request:
+    paths:
+      - bindings/go/**
+      - whisper.h
+
+jobs:
+  ubuntu-latest:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/setup-go@v3
+        with:
+          go-version: '^1.19'
+      - uses: actions/checkout@v1
+      - run: |
+          cd bindings/go
+          make test

.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

@@ -1,115 +1,308 @@
 name: CI
-on: [push]
+on: [push, pull_request]
 
 jobs:
-    ubuntu-latest:
-        runs-on: ubuntu-latest
+  ubuntu-latest:
+    runs-on: ubuntu-latest
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  sudo apt-get update
-                  sudo apt-get install build-essential
-                  sudo apt-get install libsdl2-dev
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+          sudo apt-get install libsdl2-dev
 
-            - name: Build
-              run: |
-                make
-                make stream
+      - name: Build
+        run: |
+          make
+          make stream
 
-    macOS-latest:
-        runs-on: macOS-latest
+  macOS-latest:
+    runs-on: macOS-latest
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Dependencies
-              run: |
-                  brew update
-                  brew install sdl2
+      - name: Dependencies
+        run: |
+          brew update
+          brew install sdl2
 
-            - name: Build
-              run: |
-                make
-                make stream
+      - name: Build
+        run: |
+          make
+          make stream
 
-    ubuntu-latest-gcc:
-        runs-on: ubuntu-latest
+  ubuntu-latest-gcc:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Debug, Release]
+    strategy:
+      matrix:
+        build: [Debug, Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    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: Dependencies
+        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=${{ matrix.build }}
+      - name: Configure
+        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
 
-    ubuntu-latest-clang:
-        runs-on: ubuntu-latest
+  ubuntu-latest-clang:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Debug, Release]
+    strategy:
+      matrix:
+        build: [Debug, Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    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: Dependencies
+        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=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+      - name: Configure
+        run: cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
 
-    ubuntu-latest-gcc-sanitized:
-        runs-on: ubuntu-latest
+  ubuntu-latest-gcc-sanitized:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                sanitizer: [ADDRESS, THREAD, UNDEFINED]
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    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
+      - name: Dependencies
+        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_${{ matrix.sanitizer }}=ON
+      - name: Configure
+        run: cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
 
-            - name: Build
-              run: |
-                make
-                ctest -L gh --output-on-failure
+      - name: Build
+        run: |
+          make
+          ctest -L gh --output-on-failure
+
+  windows:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [Win32, x64]
+        sdl2: [ON]
+        include:
+          - arch: Win32
+            s2arc: x86
+          - arch: x64
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
+
+      - 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_SUPPORT_SDL2=${{ matrix.sdl2 }}
+
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ 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-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
+
+  windows-blas:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [Win32, x64]
+        blas: [ON]
+        sdl2: [ON]
+        include:
+          - arch: Win32
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip
+            s2arc: x86
+          - arch: x64
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
+
+      - name: Fetch OpenBLAS
+        if: matrix.blas == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }}
+          7z x blas.zip -oblas -y
+          copy blas/include/cblas.h .
+          copy blas/include/openblas_config.h .
+          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ 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_SUPPORT_OPENBLAS=${{ matrix.blas }}
+          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
+          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}
+
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+
+      - name: Copy libopenblas.dll
+        if: matrix.blas == 'ON'
+        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+
+      - name: Upload binaries
+        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-blas-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
+
+  emscripten:
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - 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: Configure
+        run: echo "tmp"
+
+      - 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@v1
+
+      - name: Configure
+        run: cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
+
+      - 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@v1
+
+      - 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

.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

@@ -1,5 +1,8 @@
 *.o
+*.a
 .cache/
+.coreml/
+.test/
 .vs/
 .vscode/
 .DS_Store
@@ -8,15 +11,31 @@ build/
 build-em/
 build-debug/
 build-release/
+build-static/
+build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
 
-main
-stream
-bench
+/main
+/stream
+/command
+/talk
+/talk-llama
+/bench
+
+arm_neon.h
 sync.sh
+libwhisper.a
+libwhisper.so
 compile_commands.json
 
+examples/arm_neon.h
 examples/whisper.objc/whisper.objc.xcodeproj/xcshareddata
 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

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "bindings/ios"]
+	path = bindings/ios
+	url = https://github.com/ggerganov/whisper.spm

CMakeLists.txt

@@ -1,14 +1,26 @@
 cmake_minimum_required (VERSION 3.0)
-project(whisper.cpp VERSION 1.0.0)
 
-set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
+project(whisper.cpp VERSION 1.3.0)
+
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+    add_compile_options(/utf-8)
+endif ()
+
+# Add path to modules
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
-set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")
 
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(WHISPER_STANDALONE ON)
-    include(cmake/GitVars.cmake)
-    include(cmake/BuildTypes.cmake)
+    include(GitVars)
+    include(BuildTypes)
+
+    # configure project version
+    if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
+        configure_file(${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl ${CMAKE_SOURCE_DIR}/bindings/ios/Makefile @ONLY)
+    endif()
+    configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
 else()
     set(WHISPER_STANDALONE OFF)
 endif()
@@ -43,6 +55,11 @@ 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)
+
+    option(WHISPER_COREML              "whisper: enable Core ML framework" OFF)
 else()
     option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
 endif()
@@ -73,26 +90,42 @@ endif()
 
 # dependencies
 
-set(CMAKE_C_STANDARD   11)
-set(CMAKE_CXX_STANDARD 20)
-
 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_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()
     endif()
 endif()
 
 if (WHISPER_SUPPORT_OPENBLAS)
-    find_library(OPENBLAS_LIB openblas)
+    find_library(OPENBLAS_LIB
+        NAMES openblas libopenblas
+        )
     if (OPENBLAS_LIB)
         message(STATUS "OpenBLAS found")
 
@@ -120,6 +153,13 @@ if (WHISPER_ALL_WARNINGS)
             -Wcast-qual                     \
             -Wstrict-prototypes             \
             -Wpointer-arith                 \
+            -Wno-unused-function            \
+        ")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
+            -Wall                           \
+            -Wextra                         \
+            -Wpedantic                      \
+            -Wcast-qual                     \
         ")
     else()
         # todo : msvc
@@ -138,15 +178,26 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES
 else()
     message(STATUS "x86 detected")
     if (MSVC)
+        set(CMAKE_CXX_FLAGS "${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_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /arch:AVX2")
     else()
         if (EMSCRIPTEN)
-            # we require support for WASM SIMD 128-bit
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -msimd128")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
             set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
         else()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx -mavx2 -mfma -mf16c")
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
+            endif()
+            if(NOT WHISPER_NO_AVX2)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
+            endif()
+            if(NOT WHISPER_NO_FMA)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
+            endif()
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
         endif()
     endif()
 endif()
@@ -155,6 +206,33 @@ 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()
+
 #
 # whisper - this is the main library of the project
 #
@@ -162,14 +240,22 @@ endif()
 set(TARGET whisper)
 
 add_library(${TARGET}
+    ggml.h
     ggml.c
+    whisper.h
     whisper.cpp
     )
 
+include(DefaultTargetOptions)
+
 target_include_directories(${TARGET} PUBLIC
     .
     )
 
+if (WHISPER_COREML)
+    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 
@@ -188,13 +274,21 @@ if (BUILD_SHARED_LIBS)
         )
 endif()
 
+if (EMSCRIPTEN)
+    set_target_properties(${TARGET} PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()
+
 target_compile_definitions(${TARGET} PUBLIC
     ${WHISPER_EXTRA_FLAGS}
     )
 
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")
+
 install(TARGETS ${TARGET}
     LIBRARY DESTINATION lib
     ARCHIVE DESTINATION lib/static
+    RUNTIME DESTINATION bin
+    PUBLIC_HEADER DESTINATION include
     )
 
 #
@@ -207,13 +301,11 @@ add_subdirectory(bindings)
 # programs, examples and tests
 #
 
-if (WHISPER_STANDALONE)
-    if (WHISPER_BUILD_TESTS)
-        enable_testing()
-        add_subdirectory(tests)
-    endif ()
-
-    if (WHISPER_BUILD_EXAMPLES)
-        add_subdirectory(examples)
-    endif()
+if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+    enable_testing()
+    add_subdirectory(tests)
 endif ()
+
+if (WHISPER_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+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

@@ -10,6 +10,9 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
+CCV := $(shell $(CC) --version | head -n 1)
+CXXV := $(shell $(CXX) --version | head -n 1)
+
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@@ -27,10 +30,16 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11  
-CXXFLAGS = -I. -I./examples -O3 -std=c++11
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq ($(wildcard /usr/include/musl/*),)
+	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+endif
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -45,20 +54,85 @@ ifeq ($(UNAME_S),FreeBSD)
 	CFLAGS   += -pthread
 	CXXFLAGS += -pthread
 endif
+ifeq ($(UNAME_S),Haiku)
+	CFLAGS   += -pthread
+	CXXFLAGS += -pthread
+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 ($(UNAME_M),x86_64)
-	# AVX 512
-	CFLAGS += -mavx512f -mfma -mf16c
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+	ifeq ($(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,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+	else ifeq ($(UNAME_S),Linux)
+		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
+		ifneq (,$(findstring avx2,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+		FMA_M := $(shell grep "fma " /proc/cpuinfo)
+		ifneq (,$(findstring fma,$(FMA_M)))
+			CFLAGS += -mfma
+		endif
+		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
+		ifneq (,$(findstring f16c,$(F16C_M)))
+			CFLAGS += -mf16c
 
-	# AVX 256
-	#CFLAGS += -mavx -mavx2 -mfma -mf16c
+			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
+		endif
+		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
+		ifneq (,$(findstring sse3,$(SSE3_M)))
+			CFLAGS += -msse3
+		endif
+	else ifeq ($(UNAME_S),Haiku)
+		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
+		ifneq (,$(findstring avx2,$(AVX2_M)))
+			CFLAGS += -mavx2
+		endif
+		FMA_M := $(shell sysinfo -cpu | grep "FMA ")
+		ifneq (,$(findstring fma,$(FMA_M)))
+			CFLAGS += -mfma
+		endif
+		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
+		ifneq (,$(findstring f16c,$(F16C_M)))
+			CFLAGS += -mf16c
+
+			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
+		endif
+	else
+		CFLAGS += -mfma -mf16c -mavx -mavx2
+	endif
 endif
 ifeq ($(UNAME_M),amd64)
 	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
+ifneq ($(filter ppc64%,$(UNAME_M)),)
+	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
+	ifneq (,$(findstring POWER9,$(POWER9_M)))
+		CFLAGS += -mpower9-vector
+	endif
+	# Require c++23's std::byteswap for big-endian support.
+	ifeq ($(UNAME_M),ppc64)
+		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
+	endif
+endif
 ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
@@ -66,15 +140,32 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
+ifdef WHISPER_COREML
+	CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS  += -framework Foundation -framework CoreML
+endif
+ifdef WHISPER_OPENBLAS
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	LDFLAGS += -lopenblas
+endif
+ifdef WHISPER_GPROF
+	CFLAGS   += -pg
+	CXXFLAGS += -pg
+endif
 ifneq ($(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%,$(UNAME_M)),)
-	# Raspberry Pi 4
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# 32-bit ARM, for example on Armbian or possibly raspbian
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+
+	# 64-bit ARM, use these (TODO: auto-detect 64-bit)
+	# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
@@ -82,24 +173,52 @@ ifneq ($(filter armv8%,$(UNAME_M)),)
 endif
 
 #
-# Build library + main
+# Print build information
 #
 
-main: examples/main/main.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp whisper.o ggml.o -o main $(LDFLAGS)
-	./main -h
+$(info I whisper.cpp build info: )
+$(info I UNAME_S:  $(UNAME_S))
+$(info I UNAME_P:  $(UNAME_P))
+$(info I UNAME_M:  $(UNAME_M))
+$(info I CFLAGS:   $(CFLAGS))
+$(info I CXXFLAGS: $(CXXFLAGS))
+$(info I LDFLAGS:  $(LDFLAGS))
+$(info I CC:       $(CCV))
+$(info I CXX:      $(CXXV))
+$(info )
+
+default: main bench
+
+#
+# Build library
+#
 
 ggml.o: ggml.c ggml.h
 	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
 
-whisper.o: whisper.cpp whisper.h
+whisper.o: whisper.cpp whisper.h ggml.h
 	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
 
-libwhisper.a: ggml.o whisper.o
-	$(AR) rcs libwhisper.a ggml.o whisper.o
+ifndef WHISPER_COREML
+WHISPER_OBJ = whisper.o
+else
+whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -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
+
+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 bench libwhisper.a
+	rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -107,11 +226,27 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-stream: examples/stream/stream.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+SRC_COMMON = examples/common.cpp
+SRC_COMMON_SDL = examples/common-sdl.cpp
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+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
+
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(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)
+
+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)
+
+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
@@ -148,9 +283,10 @@ samples:
 .PHONY: small
 .PHONY: medium.en
 .PHONY: medium
+.PHONY: large-v1
 .PHONY: large
 
-tiny.en tiny base.en base small.en small medium.en medium large: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="
@@ -159,9 +295,17 @@ tiny.en tiny base.en base small.en small medium.en medium large: main
 	@echo ""
 	@for f in samples/*.wav; do \
 		echo "----------------------------------------------" ; \
-		echo "[+] Running base.en on $$f ... (run 'ffplay $$f' to listen)" ; \
+		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
 	    echo "----------------------------------------------" ; \
 		echo "" ; \
 		./main -m models/ggml-$@.bin -f $$f ; \
 		echo "" ; \
 	done
+
+#
+# Tests
+#
+
+.PHONY: tests
+tests:
+	bash ./tests/run-tests.sh

README.md

@@ -2,14 +2,18 @@
 
 [![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/)
+
+Beta: [v1.3.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.3.0) / 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 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
-- Low memory usage (Flash Attention + Flash Forward)
+- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
 - Runs on the CPU
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
@@ -18,11 +22,11 @@ Supported platforms:
 
 - [x] Mac OS (Intel and Arm)
 - [x] [iOS](examples/whisper.objc)
-- [x] Linux
+- [x] [Android](examples/whisper.android)
+- [x] Linux / [FreeBSD](https://github.com/ggerganov/whisper.cpp/issues/56#issuecomment-1350920264)
 - [x] [WebAssembly](examples/whisper.wasm)
-- [x] [Windows (MSVC and MinGW)](https://github.com/ggerganov/whisper.cpp/issues/5)
-- [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/issues/7)
-- [x] [Android](https://github.com/ggerganov/whisper.cpp/issues/30)
+- [x] Windows ([MSVC](https://github.com/ggerganov/whisper.cpp/blob/master/.github/workflows/build.yml#L117-L144) and [MinGW](https://github.com/ggerganov/whisper.cpp/issues/168)]
+- [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/discussions/166)
 
 The entire implementation of the model is contained in 2 source files:
 
@@ -30,10 +34,16 @@ The entire implementation of the model is contained in 2 source files:
 - Transformer inference: [whisper.h](whisper.h) / [whisper.cpp](whisper.cpp)
 
 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
-As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device:
+As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device: [whisper.objc](examples/whisper.objc)
 
 https://user-images.githubusercontent.com/1991296/197385372-962a6dea-bca1-4d50-bf96-1d8c27b98c81.mp4
 
+You can also easily make your own offline voice assistant application: [command](examples/command)
+
+https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
+
+Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
+
 ## Implementation details
 
 - The core tensor operations are implemented in C ([ggml.h](ggml.h) / [ggml.c](ggml.c))
@@ -46,24 +56,11 @@ The tensor operators are optimized heavily for Apple silicon CPUs. Depending on
 instrisics 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.
 
-## Limitations
-
-- Inference only
-- No GPU support
-- Very basic greedy sampling scheme - always pick up the token with highest probability.
-  This should be similar to the [GreedyDecoder](https://github.com/openai/whisper/blob/main/whisper/decoding.py#L249-L274)
-  from the original python implementation, so in order to make a fair comparison between the 2 implementations, make sure
-  to run the python code with the following parameters:
-
-  ```
-  whisper --best_of None --beam_size None ...
-  ```
-
-  In the future, `whisper.cpp` will support more sampling strategies.
-
 ## 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
@@ -76,7 +73,7 @@ Now build the [main](examples/main) example and transcribe an audio file like th
 make
 
 # transcribe an audio file
-./main -f input.wav
+./main -f samples/jfk.wav
 ```
 
 ---
@@ -94,27 +91,38 @@ c++ -I. -I./examples -O3 -std=c++11 -pthread examples/main/main.cpp whisper.o gg
 usage: ./main [options] file0.wav file1.wav ...
 
 options:
-  -h,       --help           show this help message and exit
-  -s SEED,  --seed SEED      RNG seed (default: -1)
-  -t N,     --threads N      number of threads to use during computation (default: 4)
-  -p N,     --processors N   number of processors to use during computation (default: 1)
-  -ot N,    --offset-t N     time offset in milliseconds (default: 0)
-  -on N,    --offset-n N     segment index offset (default: 0)
-  -mc N,    --max-context N  maximum number of text context tokens to store (default: max)
-  -ml N,    --max-len N      maximum segment length in characters (default: 0)
-  -wt N,    --word-thold N   word timestamp probability threshold (default: 0.010000)
-  -v,       --verbose        verbose output
-            --translate      translate from source language to english
-  -otxt,    --output-txt     output result in a text file
-  -ovtt,    --output-vtt     output result in a vtt file
-  -osrt,    --output-srt     output result in a srt file
-  -owts,    --output-words   output script for generating karaoke video
-  -ps,      --print_special  print special tokens
-  -pc,      --print_colors   print colors
-  -nt,      --no_timestamps  do not print timestamps
-  -l LANG,  --language LANG  spoken language (default: en)
-  -m FNAME, --model FNAME    model path (default: models/ggml-base.en.bin)
-  -f FNAME, --file FNAME     input WAV file path
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -tr,       --translate         [false  ] translate from source language to english
+  -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
+  -ovtt,     --output-vtt        [false  ] output result in a vtt file
+  -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
+  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -f FNAME,  --file FNAME        [       ] input WAV file path
+
 
 bash ./models/download-ggml-model.sh base.en
 Downloading ggml model base.en ...
@@ -133,7 +141,8 @@ Running base.en on all samples in ./samples ...
 [+] Running base.en on samples/jfk.wav ... (run 'ffplay samples/jfk.wav' to listen)
 ----------------------------------------------
 
-whisper_model_load: loading model from 'models/ggml-base.en.bin'
+whisper_init_from_file: loading model from 'models/ggml-base.en.bin'
+whisper_model_load: loading model
 whisper_model_load: n_vocab       = 51864
 whisper_model_load: n_audio_ctx   = 1500
 whisper_model_load: n_audio_state = 512
@@ -146,13 +155,14 @@ whisper_model_load: n_text_layer  = 6
 whisper_model_load: n_mels        = 80
 whisper_model_load: f16           = 1
 whisper_model_load: type          = 2
-whisper_model_load: mem_required  = 670.00 MB
+whisper_model_load: mem required  =  215.00 MB (+    6.00 MB per decoder)
+whisper_model_load: kv self size  =    5.25 MB
+whisper_model_load: kv cross size =   17.58 MB
 whisper_model_load: adding 1607 extra tokens
-whisper_model_load: ggml ctx size = 140.60 MB
-whisper_model_load: memory size =    22.83 MB
-whisper_model_load: model size  =   140.54 MB
+whisper_model_load: model ctx     =  140.60 MB
+whisper_model_load: model size    =  140.54 MB
 
-system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
+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 |
 
 main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -160,12 +170,13 @@ main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 proc
 [00:00:00.000 --> 00:00:11.000]   And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
 
 
-whisper_print_timings:     load time =   105.91 ms
-whisper_print_timings:      mel time =    24.62 ms
-whisper_print_timings:   sample time =     3.63 ms
-whisper_print_timings:   encode time =   324.71 ms / 54.12 ms per layer
-whisper_print_timings:   decode time =    83.58 ms / 13.93 ms per layer
-whisper_print_timings:    total time =   542.81 ms
+whisper_print_timings:     fallbacks =   0 p /   0 h
+whisper_print_timings:     load time =   113.81 ms
+whisper_print_timings:      mel time =    15.40 ms
+whisper_print_timings:   sample time =    11.58 ms /    27 runs (    0.43 ms per run)
+whisper_print_timings:   encode time =   266.60 ms /     1 runs (  266.60 ms per run)
+whisper_print_timings:   decode time =    66.11 ms /    27 runs (    2.45 ms per run)
+whisper_print_timings:    total time =   476.31 ms
 ```
 
 The command downloads the `base.en` model converted to custom `ggml` format and runs the inference on all `.wav` samples in the folder `samples`.
@@ -200,6 +211,7 @@ make small.en
 make small
 make medium.en
 make medium
+make large-v1
 make large
 ```
 
@@ -207,11 +219,70 @@ make large
 
 | Model  | Disk   | Mem     | SHA                                        |
 | ---    | ---    | ---     | ---                                        |
-| tiny   |  75 MB | ~390 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
-| base   | 142 MB | ~500 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
-| small  | 466 MB | ~1.0 GB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
-| medium | 1.5 GB | ~2.6 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
-| large  | 2.9 GB | ~4.7 GB | `b1caaf735c4cc1429223d5a74f0f4d0b9b59a299` |
+| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
+| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
+| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
+| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
+| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+
+## 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
+  ```
+
+- 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
+  cd build
+  cmake -DWHISPER_COREML=1 ..
+  ```
+
+- 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).
+  
+## Limitations
+
+- Inference only
+- No GPU support (yet)
 
 ## Another example
 
@@ -224,7 +295,8 @@ in about half a minute on a MacBook M1 Pro, using `medium.en` model:
 ```java
 $ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8
 
-whisper_model_load: loading model from 'models/ggml-medium.en.bin'
+whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
+whisper_model_load: loading model
 whisper_model_load: n_vocab       = 51864
 whisper_model_load: n_audio_ctx   = 1500
 whisper_model_load: n_audio_state = 1024
@@ -237,65 +309,71 @@ whisper_model_load: n_text_layer  = 24
 whisper_model_load: n_mels        = 80
 whisper_model_load: f16           = 1
 whisper_model_load: type          = 4
-whisper_model_load: mem_required  = 2610.00 MB
+whisper_model_load: mem required  = 1720.00 MB (+   43.00 MB per decoder)
+whisper_model_load: kv self size  =   42.00 MB
+whisper_model_load: kv cross size =  140.62 MB
 whisper_model_load: adding 1607 extra tokens
-whisper_model_load: ggml ctx size = 1644.97 MB
-whisper_model_load: memory size =   182.62 MB
-whisper_model_load: model size  =  1462.12 MB
+whisper_model_load: model ctx     = 1462.35 MB
+whisper_model_load: model size    = 1462.12 MB
 
-main: processing 'samples/gb1.wav' (3179750 samples, 198.7 sec), 8 threads, lang = en, task = transcribe, timestamps = 1 ...
+system_info: n_threads = 8 / 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 |
 
-[00:00.000 --> 00:08.000]   My fellow Americans, this day has brought terrible news and great sadness to our country.
-[00:08.000 --> 00:17.000]   At nine o'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.
-[00:17.000 --> 00:23.000]   A short time later, debris was seen falling from the skies above Texas.
-[00:23.000 --> 00:29.000]   The Columbia's lost. There are no survivors.
-[00:29.000 --> 00:32.000]   On board was a crew of seven.
-[00:32.000 --> 00:39.000]   Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark,
-[00:39.000 --> 00:48.000]   Captain David Brown, Commander William McCool, Dr. Kultna Shavla, and Ilan Ramon,
-[00:48.000 --> 00:52.000]   a colonel in the Israeli Air Force.
-[00:52.000 --> 00:58.000]   These men and women assumed great risk in the service to all humanity.
-[00:58.000 --> 01:03.000]   In an age when space flight has come to seem almost routine,
-[01:03.000 --> 01:07.000]   it is easy to overlook the dangers of travel by rocket
-[01:07.000 --> 01:12.000]   and the difficulties of navigating the fierce outer atmosphere of the Earth.
-[01:12.000 --> 01:18.000]   These astronauts knew the dangers, and they faced them willingly,
-[01:18.000 --> 01:23.000]   knowing they had a high and noble purpose in life.
-[01:23.000 --> 01:31.000]   Because of their courage and daring and idealism, we will miss them all the more.
-[01:31.000 --> 01:36.000]   All Americans today are thinking as well of the families of these men and women
-[01:36.000 --> 01:40.000]   who have been given this sudden shock and grief.
-[01:40.000 --> 01:45.000]   You're not alone. Our entire nation grieves with you,
-[01:45.000 --> 01:52.000]   and those you love will always have the respect and gratitude of this country.
-[01:52.000 --> 01:56.000]   The cause in which they died will continue.
-[01:56.000 --> 02:04.000]   Mankind is led into the darkness beyond our world by the inspiration of discovery
-[02:04.000 --> 02:11.000]   and the longing to understand. Our journey into space will go on.
-[02:11.000 --> 02:16.000]   In the skies today, we saw destruction and tragedy.
-[02:16.000 --> 02:22.000]   Yet farther than we can see, there is comfort and hope.
-[02:22.000 --> 02:29.000]   In the words of the prophet Isaiah, "Lift your eyes and look to the heavens
-[02:29.000 --> 02:35.000]   who created all these. He who brings out the starry hosts one by one
-[02:35.000 --> 02:39.000]   and calls them each by name."
-[02:39.000 --> 02:46.000]   Because of His great power and mighty strength, not one of them is missing.
-[02:46.000 --> 02:55.000]   The same Creator who names the stars also knows the names of the seven souls we mourn today.
-[02:55.000 --> 03:01.000]   The crew of the shuttle Columbia did not return safely to earth,
-[03:01.000 --> 03:05.000]   yet we can pray that all are safely home.
-[03:05.000 --> 03:13.000]   May God bless the grieving families, and may God continue to bless America.
-[03:13.000 --> 03:41.000]   Audio
+main: processing 'samples/gb1.wav' (3179750 samples, 198.7 sec), 8 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
 
-whisper_print_timings:     load time =   575.92 ms
-whisper_print_timings:      mel time =   230.60 ms
-whisper_print_timings:   sample time =    73.19 ms
-whisper_print_timings:   encode time = 19552.61 ms / 814.69 ms per layer
-whisper_print_timings:   decode time = 13249.96 ms / 552.08 ms per layer
-whisper_print_timings:    total time = 33686.27 ms
+[00:00:00.000 --> 00:00:08.000]   My fellow Americans, this day has brought terrible news and great sadness to our country.
+[00:00:08.000 --> 00:00:17.000]   At nine o'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.
+[00:00:17.000 --> 00:00:23.000]   A short time later, debris was seen falling from the skies above Texas.
+[00:00:23.000 --> 00:00:29.000]   The Columbia's lost. There are no survivors.
+[00:00:29.000 --> 00:00:32.000]   On board was a crew of seven.
+[00:00:32.000 --> 00:00:39.000]   Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark,
+[00:00:39.000 --> 00:00:48.000]   Captain David Brown, Commander William McCool, Dr. Kultna Shavla, and Ilan Ramon,
+[00:00:48.000 --> 00:00:52.000]   a colonel in the Israeli Air Force.
+[00:00:52.000 --> 00:00:58.000]   These men and women assumed great risk in the service to all humanity.
+[00:00:58.000 --> 00:01:03.000]   In an age when space flight has come to seem almost routine,
+[00:01:03.000 --> 00:01:07.000]   it is easy to overlook the dangers of travel by rocket
+[00:01:07.000 --> 00:01:12.000]   and the difficulties of navigating the fierce outer atmosphere of the Earth.
+[00:01:12.000 --> 00:01:18.000]   These astronauts knew the dangers, and they faced them willingly,
+[00:01:18.000 --> 00:01:23.000]   knowing they had a high and noble purpose in life.
+[00:01:23.000 --> 00:01:31.000]   Because of their courage and daring and idealism, we will miss them all the more.
+[00:01:31.000 --> 00:01:36.000]   All Americans today are thinking as well of the families of these men and women
+[00:01:36.000 --> 00:01:40.000]   who have been given this sudden shock and grief.
+[00:01:40.000 --> 00:01:45.000]   You're not alone. Our entire nation grieves with you,
+[00:01:45.000 --> 00:01:52.000]   and those you love will always have the respect and gratitude of this country.
+[00:01:52.000 --> 00:01:56.000]   The cause in which they died will continue.
+[00:01:56.000 --> 00:02:04.000]   Mankind is led into the darkness beyond our world by the inspiration of discovery
+[00:02:04.000 --> 00:02:11.000]   and the longing to understand. Our journey into space will go on.
+[00:02:11.000 --> 00:02:16.000]   In the skies today, we saw destruction and tragedy.
+[00:02:16.000 --> 00:02:22.000]   Yet farther than we can see, there is comfort and hope.
+[00:02:22.000 --> 00:02:29.000]   In the words of the prophet Isaiah, "Lift your eyes and look to the heavens
+[00:02:29.000 --> 00:02:35.000]   who created all these. He who brings out the starry hosts one by one
+[00:02:35.000 --> 00:02:39.000]   and calls them each by name."
+[00:02:39.000 --> 00:02:46.000]   Because of His great power and mighty strength, not one of them is missing.
+[00:02:46.000 --> 00:02:55.000]   The same Creator who names the stars also knows the names of the seven souls we mourn today.
+[00:02:55.000 --> 00:03:01.000]   The crew of the shuttle Columbia did not return safely to earth,
+[00:03:01.000 --> 00:03:05.000]   yet we can pray that all are safely home.
+[00:03:05.000 --> 00:03:13.000]   May God bless the grieving families, and may God continue to bless America.
+[00:03:13.000 --> 00:03:19.000]   [Silence]
+
+
+whisper_print_timings:     fallbacks =   1 p /   0 h
+whisper_print_timings:     load time =   569.03 ms
+whisper_print_timings:      mel time =   146.85 ms
+whisper_print_timings:   sample time =   238.66 ms /   553 runs (    0.43 ms per run)
+whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per run)
+whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
+whisper_print_timings:    total time = 32733.52 ms
 ```
 </details>
 
 ## 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
+make stream
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
@@ -306,18 +384,22 @@ 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)
 
-For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`: 
+For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:
 
 ```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
-system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 
+system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -341,7 +423,7 @@ The `--max-len` argument can be used to obtain word-level timestamps. Simply use
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
-system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 
+system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 |
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
@@ -411,6 +493,19 @@ 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,
@@ -428,18 +523,57 @@ The original models are converted to a custom binary format. This allows to pack
 - vocabulary
 - weights
 
-You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script or from here:
+You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
+or manually from here:
 
-https://ggml.ggerganov.com
+- 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 in [models](models).
+For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
+in [models](models).
 
-## Bindings
+## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)
 
-- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs)
-- [ ] Python:
-- [ ] Java:
+- [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] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
+- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
+- [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
 
-There are various examples of using the library for different projects in the [examples](examples) folder. Check them out!
+There are various examples of using the library for different projects in the [examples](examples) folder.
+Some of the examples are even ported to run in the browser using WebAssembly. Check them out!
+
+| Example | Web | Description |
+| ---     | --- | ---         |
+| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
+| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
+| [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 |
+| [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
+| [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
+| [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
+| [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
+
+## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)
+
+If you have any kind of feedback about this project feel free to use the Discussions section and open a new topic.
+You can use the [Show and tell](https://github.com/ggerganov/whisper.cpp/discussions/categories/show-and-tell) category
+to share your own projects that use `whisper.cpp`. If you have a question, make sure to check the
+[Frequently asked questions (#126)](https://github.com/ggerganov/whisper.cpp/discussions/126) discussion.

bindings/CMakeLists.txt

@@ -1,3 +1,19 @@
 if (EMSCRIPTEN)
     add_subdirectory(javascript)
+
+    add_custom_command(
+        OUTPUT ${CMAKE_CURRENT_SOURCE_DIR}/javascript/publish.log
+        DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/javascript/whisper.js
+        DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/javascript/libwhisper.worker.js
+        DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/javascript/package.json
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/javascript
+        COMMAND npm publish
+        COMMAND touch publish.log
+        COMMENT "Publishing npm module v${PROJECT_VERSION}"
+        VERBATIM
+        )
+
+    add_custom_target(publish-npm
+        DEPENDS javascript/publish.log
+        )
 endif()

bindings/go/.gitignore

@@ -0,0 +1,2 @@
+build
+models

bindings/go/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 David Thorpe
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

bindings/go/Makefile

@@ -0,0 +1,38 @@
+BUILD_DIR := build
+MODELS_DIR := models
+EXAMPLES_DIR := $(wildcard examples/*)
+INCLUDE_PATH := $(abspath ../..)
+LIBRARY_PATH := $(abspath ../..)
+
+all: clean whisper examples
+
+whisper: mkdir
+	@echo Build whisper
+	@${MAKE} -C ../.. libwhisper.a
+
+test: model-small whisper modtidy
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v .
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v ./pkg/whisper/...
+
+examples: $(EXAMPLES_DIR)
+
+model-small: mkdir examples/go-model-download
+	@${BUILD_DIR}/go-model-download -out models ggml-small.en.bin
+
+$(EXAMPLES_DIR): mkdir whisper modtidy
+	@echo Build example $(notdir $@)
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
+
+mkdir:
+	@echo Mkdir ${BUILD_DIR}
+	@install -d ${BUILD_DIR}
+	@echo Mkdir ${MODELS_DIR}
+	@install -d ${MODELS_DIR}
+
+modtidy:
+	@go mod tidy
+
+clean: 
+	@echo Clean
+	@rm -fr $(BUILD_DIR)
+	@go clean

bindings/go/README.md

@@ -0,0 +1,100 @@
+# Go bindings for Whisper
+
+This package provides Go bindings for whisper.cpp. They have been tested on:
+
+  * Darwin (OS X) 12.6 on x64_64
+  * Debian Linux on arm64
+  * Fedora Linux on x86_64
+
+The "low level" bindings are in the `bindings/go` directory and there is a more
+Go-style package in the `bindings/go/pkg/whisper` directory. The most simple usage
+is as follows:
+
+```go
+import (
+	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+)
+
+func main() {
+	var modelpath string // Path to the model
+	var samples []float32 // Samples to process
+
+	// Load the model
+	model, err := whisper.New(modelpath)
+	if err != nil {
+		panic(err)
+	}
+	defer model.Close()
+
+	// Process samples
+	context, err := model.NewContext()
+	if err != nil {
+		panic(err)
+	}
+	if err := context.Process(samples, nil); err != nil {
+		return err
+	}
+
+	// Print out the results
+	for {
+		segment, err := context.NextSegment()
+		if err != nil {
+			break
+		}
+		fmt.Printf("[%6s->%6s] %s\n", segment.Start, segment.End, segment.Text)
+	}
+}
+```
+
+## Building & Testing
+
+In order to build, you need to have the Go compiler installed. You can get it from [here](https://golang.org/dl/). Run the tests with:
+
+```bash
+git clone https://github.com/ggerganov/whisper.cpp.git
+cd whisper.cpp/bindings/go
+make test
+```
+
+This will compile a static `libwhisper.a` in a `build` folder, download a model file, then run the tests. To build the examples:
+
+```bash
+make examples
+```
+
+The examples are placed in the `build` directory. Once built, you can download all the models with the following command:
+
+```bash
+./build/go-model-download -out models
+```
+
+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 
+```
+
+## Using the bindings
+
+To use the bindings in your own software,
+
+  1. Import `github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper` (or `github.com/ggerganov/whisper.cpp/bindings/go` into your package;
+  2. Compile `libwhisper.a` (you can use `make whisper` in the `bindings/go` directory);
+  3. Link your go binary against whisper by setting the environment variables `C_INCLUDE_PATH` and `LIBRARY_PATH`
+     to point to the `whisper.h` file directory and `libwhisper.a` file directory respectively.
+
+Look at the `Makefile` in the `bindings/go` directory for an example.
+
+The API Documentation:
+
+  * https://pkg.go.dev/github.com/ggerganov/whisper.cpp/bindings/go
+  * https://pkg.go.dev/github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper
+
+Getting help:
+
+  * Follow the discussion for the go bindings [here](https://github.com/ggerganov/whisper.cpp/discussions/312)
+
+## 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/go/doc.go

@@ -0,0 +1,5 @@
+/*
+github.com/ggerganov/whisper.cpp/bindings/go
+provides a speech-to-text service bindings for the Go programming language.
+*/
+package whisper

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

@@ -0,0 +1,30 @@
+package main
+
+import (
+	"context"
+	"os"
+	"os/signal"
+)
+
+// ContextForSignal returns a context object which is cancelled when a signal
+// is received. It returns nil if no signal parameter is provided
+func ContextForSignal(signals ...os.Signal) context.Context {
+	if len(signals) == 0 {
+		return nil
+	}
+
+	ch := make(chan os.Signal)
+	ctx, cancel := context.WithCancel(context.Background())
+
+	// Send message on channel when signal received
+	signal.Notify(ch, signals...)
+
+	// When any signal received, call cancel
+	go func() {
+		<-ch
+		cancel()
+	}()
+
+	// Return success
+	return ctx
+}

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

@@ -0,0 +1,208 @@
+package main
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"path/filepath"
+	"syscall"
+	"time"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+const (
+	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 (
+	// The models which will be downloaded, if no model is specified as an argument
+	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large"}
+)
+
+var (
+	// The output folder. When not set, use current working directory.
+	flagOut = flag.String("out", "", "Output folder")
+
+	// HTTP timeout parameter - will timeout if takes longer than this to download a model
+	flagTimeout = flag.Duration("timeout", 30*time.Minute, "HTTP timeout")
+
+	// Quiet parameter - will not print progress if set
+	flagQuiet = flag.Bool("quiet", false, "Quiet mode")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// MAIN
+
+func main() {
+	flag.Usage = func() {
+		name := filepath.Base(flag.CommandLine.Name())
+		fmt.Fprintf(flag.CommandLine.Output(), "Usage: %s [options] <model>\n\n", name)
+		flag.PrintDefaults()
+	}
+	flag.Parse()
+
+	// Get output path
+	out, err := GetOut()
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "Error:", err)
+		os.Exit(-1)
+	}
+
+	// Create context which quits on SIGINT or SIGQUIT
+	ctx := ContextForSignal(os.Interrupt, syscall.SIGQUIT)
+
+	// Progress filehandle
+	progress := os.Stdout
+	if *flagQuiet {
+		progress, err = os.Open(os.DevNull)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			os.Exit(-1)
+		}
+		defer progress.Close()
+	}
+
+	// Download models - exit on error or interrupt
+	for _, model := range GetModels() {
+		url, err := URLForModel(model)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			continue
+		} else if path, err := Download(ctx, progress, url, out); err == nil || err == io.EOF {
+			continue
+		} else if err == context.Canceled {
+			os.Remove(path)
+			fmt.Fprintln(progress, "\nInterrupted")
+			break
+		} else if err == context.DeadlineExceeded {
+			os.Remove(path)
+			fmt.Fprintln(progress, "Timeout downloading model")
+			continue
+		} else {
+			os.Remove(path)
+			fmt.Fprintln(os.Stderr, "Error:", err)
+			break
+		}
+	}
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// GetOut returns the path to the output directory
+func GetOut() (string, error) {
+	if *flagOut == "" {
+		return os.Getwd()
+	}
+	if info, err := os.Stat(*flagOut); err != nil {
+		return "", err
+	} else if !info.IsDir() {
+		return "", fmt.Errorf("not a directory: %s", info.Name())
+	} else {
+		return *flagOut, nil
+	}
+}
+
+// GetModels returns the list of models to download
+func GetModels() []string {
+	if flag.NArg() == 0 {
+		return modelNames
+	} else {
+		return flag.Args()
+	}
+}
+
+// URLForModel returns the URL for the given model on huggingface.co
+func URLForModel(model string) (string, error) {
+	if filepath.Ext(model) != srcExt {
+		model += srcExt
+	}
+	url, err := url.Parse(srcUrl)
+	if err != nil {
+		return "", err
+	} else {
+		url.Path = filepath.Join(url.Path, model)
+	}
+	return url.String(), nil
+}
+
+// Download downloads the model from the given URL to the given output directory
+func Download(ctx context.Context, p io.Writer, model, out string) (string, error) {
+	// Create HTTP client
+	client := http.Client{
+		Timeout: *flagTimeout,
+	}
+
+	// Initiate the download
+	req, err := http.NewRequest("GET", model, nil)
+	if err != nil {
+		return "", err
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("%s: %s", model, resp.Status)
+	}
+
+	// If output file exists and is the same size as the model, skip
+	path := filepath.Join(out, filepath.Base(model))
+	if info, err := os.Stat(path); err == nil && info.Size() == resp.ContentLength {
+		fmt.Fprintln(p, "Skipping", model, "as it already exists")
+		return "", nil
+	}
+
+	// Create file
+	w, err := os.Create(path)
+	if err != nil {
+		return "", err
+	}
+	defer w.Close()
+
+	// Report
+	fmt.Fprintln(p, "Downloading", model, "to", out)
+
+	// Progressively download the model
+	data := make([]byte, bufSize)
+	count, pct := int64(0), int64(0)
+	ticker := time.NewTicker(5 * time.Second)
+	for {
+		select {
+		case <-ctx.Done():
+			// Cancelled, return error
+			return path, ctx.Err()
+		case <-ticker.C:
+			pct = DownloadReport(p, pct, count, resp.ContentLength)
+		default:
+			// Read body
+			n, err := resp.Body.Read(data)
+			if err != nil {
+				DownloadReport(p, pct, count, resp.ContentLength)
+				return path, err
+			} else if m, err := w.Write(data[:n]); err != nil {
+				return path, err
+			} else {
+				count += int64(m)
+			}
+		}
+	}
+}
+
+// Report periodically reports the download progress when percentage changes
+func DownloadReport(w io.Writer, pct, count, total int64) int64 {
+	pct_ := count * 100 / total
+	if pct_ > pct {
+		fmt.Fprintf(w, "  ...%d MB written (%d%%)\n", count/1e6, pct_)
+	}
+	return pct_
+}

bindings/go/examples/go-whisper/color.go

@@ -0,0 +1,22 @@
+package main
+
+import "fmt"
+
+///////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+const (
+	Reset     = "\033[0m"
+	RGBPrefix = "\033[38;5;" // followed by RGB values in decimal format separated by colons
+	RGBSuffix = "m"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Colorize text with RGB values, from 0 to 23
+func Colorize(text string, v int) string {
+	// https://en.wikipedia.org/wiki/ANSI_escape_code#8-bit
+	// Grayscale colors are in the range 232-255
+	return RGBPrefix + fmt.Sprint(v%24+232) + RGBSuffix + text + Reset
+}

bindings/go/examples/go-whisper/flags.go

@@ -0,0 +1,156 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"strings"
+	"time"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type Flags struct {
+	*flag.FlagSet
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func NewFlags(name string, args []string) (*Flags, error) {
+	flags := &Flags{
+		FlagSet: flag.NewFlagSet(name, flag.ContinueOnError),
+	}
+
+	// Register the command line arguments
+	registerFlags(flags)
+
+	// Parse command line
+	if err := flags.Parse(args); err != nil {
+		return nil, err
+	}
+
+	// Return success
+	return flags, nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+func (flags *Flags) GetModel() string {
+	return flags.Lookup("model").Value.String()
+}
+
+func (flags *Flags) GetLanguage() string {
+	return flags.Lookup("language").Value.String()
+}
+
+func (flags *Flags) IsTranslate() bool {
+	return flags.Lookup("translate").Value.(flag.Getter).Get().(bool)
+}
+
+func (flags *Flags) GetOffset() time.Duration {
+	return flags.Lookup("offset").Value.(flag.Getter).Get().(time.Duration)
+}
+
+func (flags *Flags) GetDuration() time.Duration {
+	return flags.Lookup("duration").Value.(flag.Getter).Get().(time.Duration)
+}
+
+func (flags *Flags) GetThreads() uint {
+	return flags.Lookup("threads").Value.(flag.Getter).Get().(uint)
+}
+
+func (flags *Flags) GetOut() string {
+	return strings.ToLower(flags.Lookup("out").Value.String())
+}
+
+func (flags *Flags) IsSpeedup() bool {
+	return flags.Lookup("speedup").Value.String() == "true"
+}
+
+func (flags *Flags) IsTokens() bool {
+	return flags.Lookup("tokens").Value.String() == "true"
+}
+
+func (flags *Flags) IsColorize() bool {
+	return flags.Lookup("colorize").Value.String() == "true"
+}
+
+func (flags *Flags) GetMaxLen() uint {
+	return flags.Lookup("max-len").Value.(flag.Getter).Get().(uint)
+}
+
+func (flags *Flags) GetMaxTokens() uint {
+	return flags.Lookup("max-tokens").Value.(flag.Getter).Get().(uint)
+}
+
+func (flags *Flags) GetWordThreshold() float32 {
+	return float32(flags.Lookup("word-thold").Value.(flag.Getter).Get().(float64))
+}
+
+func (flags *Flags) SetParams(context whisper.Context) error {
+	if lang := flags.GetLanguage(); lang != "" && lang != "auto" {
+		fmt.Fprintf(flags.Output(), "Setting language to %q\n", lang)
+		if err := context.SetLanguage(lang); err != nil {
+			return err
+		}
+	}
+	if flags.IsTranslate() && context.IsMultilingual() {
+		fmt.Fprintf(flags.Output(), "Setting translate to true\n")
+		context.SetTranslate(true)
+	}
+	if offset := flags.GetOffset(); offset != 0 {
+		fmt.Fprintf(flags.Output(), "Setting offset to %v\n", offset)
+		context.SetOffset(offset)
+	}
+	if duration := flags.GetDuration(); duration != 0 {
+		fmt.Fprintf(flags.Output(), "Setting duration to %v\n", duration)
+		context.SetDuration(duration)
+	}
+	if flags.IsSpeedup() {
+		fmt.Fprintf(flags.Output(), "Setting speedup to true\n")
+		context.SetSpeedup(true)
+	}
+	if threads := flags.GetThreads(); threads != 0 {
+		fmt.Fprintf(flags.Output(), "Setting threads to %d\n", threads)
+		context.SetThreads(threads)
+	}
+	if max_len := flags.GetMaxLen(); max_len != 0 {
+		fmt.Fprintf(flags.Output(), "Setting max_segment_length to %d\n", max_len)
+		context.SetMaxSegmentLength(max_len)
+	}
+	if max_tokens := flags.GetMaxTokens(); max_tokens != 0 {
+		fmt.Fprintf(flags.Output(), "Setting max_tokens to %d\n", max_tokens)
+		context.SetMaxTokensPerSegment(max_tokens)
+	}
+	if word_threshold := flags.GetWordThreshold(); word_threshold != 0 {
+		fmt.Fprintf(flags.Output(), "Setting word_threshold to %f\n", word_threshold)
+		context.SetTokenThreshold(word_threshold)
+	}
+
+	// Return success
+	return nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func registerFlags(flag *Flags) {
+	flag.String("model", "", "Path to the model file")
+	flag.String("language", "", "Spoken language")
+	flag.Bool("translate", false, "Translate from source language to english")
+	flag.Duration("offset", 0, "Time offset")
+	flag.Duration("duration", 0, "Duration of audio to process")
+	flag.Uint("threads", 0, "Number of threads to use")
+	flag.Bool("speedup", false, "Enable speedup")
+	flag.Uint("max-len", 0, "Maximum segment length in characters")
+	flag.Uint("max-tokens", 0, "Maximum tokens per segment")
+	flag.Float64("word-thold", 0, "Maximum segment score")
+	flag.Bool("tokens", false, "Display tokens")
+	flag.Bool("colorize", false, "Colorize tokens")
+	flag.String("out", "", "Output format (srt, none or leave as empty string)")
+}

bindings/go/examples/go-whisper/main.go

@@ -0,0 +1,43 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+)
+
+func main() {
+	flags, err := NewFlags(filepath.Base(os.Args[0]), os.Args[1:])
+	if err == flag.ErrHelp {
+		os.Exit(0)
+	} else if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	} else if flags.GetModel() == "" {
+		fmt.Fprintln(os.Stderr, "Use -model flag to specify which model file to use")
+		os.Exit(1)
+	} else if flags.NArg() == 0 {
+		fmt.Fprintln(os.Stderr, "No input files specified")
+		os.Exit(1)
+	}
+
+	// Load model
+	model, err := whisper.New(flags.GetModel())
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	defer model.Close()
+
+	// Process files
+	for _, filename := range flags.Args() {
+		if err := Process(model, filename, flags); err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			continue
+		}
+	}
+}

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

@@ -0,0 +1,132 @@
+package main
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"time"
+
+	// Package imports
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	wav "github.com/go-audio/wav"
+)
+
+func Process(model whisper.Model, path string, flags *Flags) error {
+	var data []float32
+
+	// Create processing context
+	context, err := model.NewContext()
+	if err != nil {
+		return err
+	}
+
+	// Set the parameters
+	if err := flags.SetParams(context); err != nil {
+		return err
+	}
+
+	fmt.Printf("\n%s\n", context.SystemInfo())
+
+	// Open the file
+	fmt.Fprintf(flags.Output(), "Loading %q\n", path)
+	fh, err := os.Open(path)
+	if err != nil {
+		return err
+	}
+	defer fh.Close()
+
+	// Decode the WAV file - load the full buffer
+	dec := wav.NewDecoder(fh)
+	if buf, err := dec.FullPCMBuffer(); err != nil {
+		return err
+	} else if dec.SampleRate != whisper.SampleRate {
+		return fmt.Errorf("unsupported sample rate: %d", dec.SampleRate)
+	} else if dec.NumChans != 1 {
+		return fmt.Errorf("unsupported number of channels: %d", dec.NumChans)
+	} else {
+		data = buf.AsFloat32Buffer().Data
+	}
+
+	// Segment callback when -tokens is specified
+	var cb whisper.SegmentCallback
+	if flags.IsTokens() {
+		cb = func(segment whisper.Segment) {
+			fmt.Fprintf(flags.Output(), "%02d [%6s->%6s] ", segment.Num, segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond))
+			for _, token := range segment.Tokens {
+				if flags.IsColorize() && context.IsText(token) {
+					fmt.Fprint(flags.Output(), Colorize(token.Text, int(token.P*24.0)), " ")
+				} else {
+					fmt.Fprint(flags.Output(), token.Text, " ")
+				}
+			}
+			fmt.Fprintln(flags.Output(), "")
+			fmt.Fprintln(flags.Output(), "")
+		}
+	}
+
+	// Process the data
+	fmt.Fprintf(flags.Output(), "  ...processing %q\n", path)
+	context.ResetTimings()
+	if err := context.Process(data, cb); err != nil {
+		return err
+	}
+
+	context.PrintTimings()
+
+	// Print out the results
+	switch {
+	case flags.GetOut() == "srt":
+		return OutputSRT(os.Stdout, context)
+	case flags.GetOut() == "none":
+		return nil
+	default:
+		return Output(os.Stdout, context, flags.IsColorize())
+	}
+}
+
+// Output text as SRT file
+func OutputSRT(w io.Writer, context whisper.Context) error {
+	n := 1
+	for {
+		segment, err := context.NextSegment()
+		if err == io.EOF {
+			return nil
+		} else if err != nil {
+			return err
+		}
+		fmt.Fprintln(w, n)
+		fmt.Fprintln(w, srtTimestamp(segment.Start), " --> ", srtTimestamp(segment.End))
+		fmt.Fprintln(w, segment.Text)
+		fmt.Fprintln(w, "")
+		n++
+	}
+}
+
+// Output text to terminal
+func Output(w io.Writer, context whisper.Context, colorize bool) error {
+	for {
+		segment, err := context.NextSegment()
+		if err == io.EOF {
+			return nil
+		} else if err != nil {
+			return err
+		}
+		fmt.Fprintf(w, "[%6s->%6s]", segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond))
+		if colorize {
+			for _, token := range segment.Tokens {
+				if !context.IsText(token) {
+					continue
+				}
+				fmt.Fprint(w, " ", Colorize(token.Text, int(token.P*24.0)))
+			}
+			fmt.Fprint(w, "\n")
+		} else {
+			fmt.Fprintln(w, " ", segment.Text)
+		}
+	}
+}
+
+// Return srtTimestamp
+func srtTimestamp(t time.Duration) string {
+	return fmt.Sprintf("%02d:%02d:%02d,%03d", t/time.Hour, (t%time.Hour)/time.Minute, (t%time.Minute)/time.Second, (t%time.Second)/time.Millisecond)
+}

bindings/go/go.mod

@@ -0,0 +1,16 @@
+module github.com/ggerganov/whisper.cpp/bindings/go
+
+go 1.19
+
+require (
+	github.com/go-audio/wav v1.1.0
+	github.com/stretchr/testify v1.8.1
+)
+
+require (
+	github.com/davecgh/go-spew v1.1.1 // indirect
+	github.com/go-audio/audio v1.0.0 // indirect
+	github.com/go-audio/riff v1.0.0 // indirect
+	github.com/pmezard/go-difflib v1.0.0 // indirect
+	gopkg.in/yaml.v3 v3.0.1 // indirect
+)

bindings/go/go.sum

@@ -0,0 +1,23 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/go-audio/audio v1.0.0 h1:zS9vebldgbQqktK4H0lUqWrG8P0NxCJVqcj7ZpNnwd4=
+github.com/go-audio/audio v1.0.0/go.mod h1:6uAu0+H2lHkwdGsAY+j2wHPNPpPoeg5AaEFh9FlA+Zs=
+github.com/go-audio/riff v1.0.0 h1:d8iCGbDvox9BfLagY94fBynxSPHO80LmZCaOsmKxokA=
+github.com/go-audio/riff v1.0.0/go.mod h1:l3cQwc85y79NQFCRB7TiPoNiaijp6q8Z0Uv38rVG498=
+github.com/go-audio/wav v1.1.0 h1:jQgLtbqBzY7G+BM8fXF7AHUk1uHUviWS4X39d5rsL2g=
+github.com/go-audio/wav v1.1.0/go.mod h1:mpe9qfwbScEbkd8uybLuIpTgHyrISw/OTuvjUW2iGtE=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
+github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

bindings/go/params.go

@@ -0,0 +1,169 @@
+package whisper
+
+import (
+	"fmt"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CGO
+
+/*
+#include <whisper.h>
+*/
+import "C"
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+func (p *Params) SetTranslate(v bool) {
+	p.translate = toBool(v)
+}
+
+func (p *Params) SetNoContext(v bool) {
+	p.no_context = toBool(v)
+}
+
+func (p *Params) SetSingleSegment(v bool) {
+	p.single_segment = toBool(v)
+}
+
+func (p *Params) SetPrintSpecial(v bool) {
+	p.print_special = toBool(v)
+}
+
+func (p *Params) SetPrintProgress(v bool) {
+	p.print_progress = toBool(v)
+}
+
+func (p *Params) SetPrintRealtime(v bool) {
+	p.print_realtime = toBool(v)
+}
+
+func (p *Params) SetPrintTimestamps(v bool) {
+	p.print_timestamps = toBool(v)
+}
+
+func (p *Params) SetSpeedup(v bool) {
+	p.speed_up = toBool(v)
+}
+
+// Set language id
+func (p *Params) SetLanguage(lang int) error {
+	if lang == -1 {
+		p.language = nil
+		return nil
+	}
+	str := C.whisper_lang_str(C.int(lang))
+	if str == nil {
+		return ErrInvalidLanguage
+	} else {
+		p.language = str
+	}
+	return nil
+}
+
+// Get language id
+func (p *Params) Language() int {
+	if p.language == nil {
+		return -1
+	}
+	return int(C.whisper_lang_id(p.language))
+}
+
+// Threads available
+func (p *Params) Threads() int {
+	return int(p.n_threads)
+}
+
+// Set number of threads to use
+func (p *Params) SetThreads(threads int) {
+	p.n_threads = C.int(threads)
+}
+
+// Set start offset in ms
+func (p *Params) SetOffset(offset_ms int) {
+	p.offset_ms = C.int(offset_ms)
+}
+
+// Set audio duration to process in ms
+func (p *Params) SetDuration(duration_ms int) {
+	p.duration_ms = C.int(duration_ms)
+}
+
+// Set timestamp token probability threshold (~0.01)
+func (p *Params) SetTokenThreshold(t float32) {
+	p.thold_pt = C.float(t)
+}
+
+// Set timestamp token sum probability threshold (~0.01)
+func (p *Params) SetTokenSumThreshold(t float32) {
+	p.thold_ptsum = C.float(t)
+}
+
+// Set max segment length in characters
+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)
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func toBool(v bool) C.bool {
+	if v {
+		return C.bool(true)
+	}
+	return C.bool(false)
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// STRINGIFY
+
+func (p *Params) String() string {
+	str := "<whisper.params"
+	str += fmt.Sprintf(" strategy=%v", p.strategy)
+	str += fmt.Sprintf(" n_threads=%d", p.n_threads)
+	if p.language != nil {
+		str += fmt.Sprintf(" language=%s", C.GoString(p.language))
+	}
+	str += fmt.Sprintf(" n_max_text_ctx=%d", p.n_max_text_ctx)
+	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
+	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
+	if p.translate {
+		str += " translate"
+	}
+	if p.no_context {
+		str += " no_context"
+	}
+	if p.single_segment {
+		str += " single_segment"
+	}
+	if p.print_special {
+		str += " print_special"
+	}
+	if p.print_progress {
+		str += " print_progress"
+	}
+	if p.print_realtime {
+		str += " print_realtime"
+	}
+	if p.print_timestamps {
+		str += " print_timestamps"
+	}
+	if p.token_timestamps {
+		str += " token_timestamps"
+	}
+	if p.speed_up {
+		str += " speed_up"
+	}
+
+	return str + ">"
+}

bindings/go/pkg/whisper/consts.go

@@ -0,0 +1,28 @@
+package whisper
+
+import (
+	"errors"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// ERRORS
+
+var (
+	ErrUnableToLoadModel    = errors.New("unable to load model")
+	ErrInternalAppError     = errors.New("internal application error")
+	ErrProcessingFailed     = errors.New("processing failed")
+	ErrUnsupportedLanguage  = errors.New("unsupported language")
+	ErrModelNotMultilingual = errors.New("model is not multilingual")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+// SampleRate is the sample rate of the audio data.
+const SampleRate = whisper.SampleRate
+
+// SampleBits is the number of bytes per sample.
+const SampleBits = whisper.SampleBits

bindings/go/pkg/whisper/context.go

@@ -0,0 +1,299 @@
+package whisper
+
+import (
+	"fmt"
+	"io"
+	"runtime"
+	"strings"
+	"time"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type context struct {
+	n      int
+	model  *model
+	params whisper.Params
+}
+
+// Make sure context adheres to the interface
+var _ Context = (*context)(nil)
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func newContext(model *model, params whisper.Params) (Context, error) {
+	context := new(context)
+	context.model = model
+	context.params = params
+
+	// Return success
+	return context, nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Set the language to use for speech recognition.
+func (context *context) SetLanguage(lang string) error {
+	if context.model.ctx == nil {
+		return ErrInternalAppError
+	}
+	if !context.model.IsMultilingual() {
+		return ErrModelNotMultilingual
+	}
+
+	if lang == "auto" {
+		context.params.SetLanguage(-1)
+	} else if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
+		return ErrUnsupportedLanguage
+	} else if err := context.params.SetLanguage(id); err != nil {
+		return err
+	}
+	// Return success
+	return nil
+}
+
+func (context *context) IsMultilingual() bool {
+	return context.model.IsMultilingual()
+}
+
+// Get language
+func (context *context) Language() string {
+	id := context.params.Language()
+	if id == -1 {
+		return "auto"
+	}
+	return whisper.Whisper_lang_str(context.params.Language())
+}
+
+// Set translate flag
+func (context *context) SetTranslate(v bool) {
+	context.params.SetTranslate(v)
+}
+
+// Set speedup flag
+func (context *context) SetSpeedup(v bool) {
+	context.params.SetSpeedup(v)
+}
+
+// Set number of threads to use
+func (context *context) SetThreads(v uint) {
+	context.params.SetThreads(int(v))
+}
+
+// Set time offset
+func (context *context) SetOffset(v time.Duration) {
+	context.params.SetOffset(int(v.Milliseconds()))
+}
+
+// Set duration of audio to process
+func (context *context) SetDuration(v time.Duration) {
+	context.params.SetOffset(int(v.Milliseconds()))
+}
+
+// Set timestamp token probability threshold (~0.01)
+func (context *context) SetTokenThreshold(t float32) {
+	context.params.SetTokenThreshold(t)
+}
+
+// Set timestamp token sum probability threshold (~0.01)
+func (context *context) SetTokenSumThreshold(t float32) {
+	context.params.SetTokenSumThreshold(t)
+}
+
+// Set max segment length in characters
+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))
+}
+
+// ResetTimings resets the mode timings. Should be called before processing
+func (context *context) ResetTimings() {
+	context.model.ctx.Whisper_reset_timings()
+}
+
+// PrintTimings prints the model timings to stdout.
+func (context *context) PrintTimings() {
+	context.model.ctx.Whisper_print_timings()
+}
+
+// SystemInfo returns the system information
+func (context *context) SystemInfo() string {
+	return fmt.Sprintf("system_info: n_threads = %d / %d | %s\n",
+		context.params.Threads(),
+		runtime.NumCPU(),
+		whisper.Whisper_print_system_info(),
+	)
+}
+
+// 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 probabilities of all languages.
+func (context *context) WhisperLangAutoDetect(offset_ms int, n_threads int) ([]float32, error) {
+	langProbs, err := context.model.ctx.Whisper_lang_auto_detect(offset_ms, n_threads)
+	if err != nil {
+		return nil, err
+	}
+	return langProbs, nil
+}
+
+// Process new sample data and return any errors
+func (context *context) Process(data []float32, cb SegmentCallback) error {
+	if context.model.ctx == nil {
+		return ErrInternalAppError
+	}
+	// If the callback is defined then we force on single_segment mode
+	if cb != nil {
+		context.params.SetSingleSegment(true)
+	}
+
+	// We don't do parallel processing at the moment
+	processors := 0
+	if processors > 1 {
+		if err := context.model.ctx.Whisper_full_parallel(context.params, data, processors, nil, func(new int) {
+			if cb != 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))
+				}
+			}
+		}); err != nil {
+			return err
+		}
+	} else if err := context.model.ctx.Whisper_full(context.params, data, nil, func(new int) {
+		if cb != 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))
+			}
+		}
+	}); err != nil {
+		return err
+	}
+
+	// Return success
+	return nil
+}
+
+// Return the next segment of tokens
+func (context *context) NextSegment() (Segment, error) {
+	if context.model.ctx == nil {
+		return Segment{}, ErrInternalAppError
+	}
+	if context.n >= context.model.ctx.Whisper_full_n_segments() {
+		return Segment{}, io.EOF
+	}
+
+	// Populate result
+	result := toSegment(context.model.ctx, context.n)
+
+	// Increment the cursor
+	context.n++
+
+	// Return success
+	return result, nil
+}
+
+// Test for text tokens
+func (context *context) IsText(t Token) bool {
+	switch {
+	case context.IsBEG(t):
+		return false
+	case context.IsSOT(t):
+		return false
+	case whisper.Token(t.Id) >= context.model.ctx.Whisper_token_eot():
+		return false
+	case context.IsPREV(t):
+		return false
+	case context.IsSOLM(t):
+		return false
+	case context.IsNOT(t):
+		return false
+	default:
+		return true
+	}
+}
+
+// Test for "begin" token
+func (context *context) IsBEG(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_beg()
+}
+
+// Test for "start of transcription" token
+func (context *context) IsSOT(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_sot()
+}
+
+// Test for "end of transcription" token
+func (context *context) IsEOT(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_eot()
+}
+
+// Test for "start of prev" token
+func (context *context) IsPREV(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_prev()
+}
+
+// Test for "start of lm" token
+func (context *context) IsSOLM(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_solm()
+}
+
+// Test for "No timestamps" token
+func (context *context) IsNOT(t Token) bool {
+	return whisper.Token(t.Id) == context.model.ctx.Whisper_token_not()
+}
+
+// Test for token associated with a specific language
+func (context *context) IsLANG(t Token, lang string) bool {
+	if id := context.model.ctx.Whisper_lang_id(lang); id >= 0 {
+		return whisper.Token(t.Id) == context.model.ctx.Whisper_token_lang(id)
+	} else {
+		return false
+	}
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PRIVATE METHODS
+
+func toSegment(ctx *whisper.Context, n int) Segment {
+	return Segment{
+		Num:    n,
+		Text:   strings.TrimSpace(ctx.Whisper_full_get_segment_text(n)),
+		Start:  time.Duration(ctx.Whisper_full_get_segment_t0(n)) * time.Millisecond * 10,
+		End:    time.Duration(ctx.Whisper_full_get_segment_t1(n)) * time.Millisecond * 10,
+		Tokens: toTokens(ctx, n),
+	}
+}
+
+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:  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/context_test.go

@@ -0,0 +1,55 @@
+package whisper_test
+
+import (
+	"os"
+	"testing"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
+	assert "github.com/stretchr/testify/assert"
+)
+
+const (
+	ModelPath  = "../../models/ggml-tiny.bin"
+	SamplePath = "../../samples/jfk.wav"
+)
+
+func Test_Whisper_000(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Load model
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	assert.NoError(model.Close())
+
+	t.Log("languages=", model.Languages())
+}
+
+func Test_Whisper_001(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Load model
+	model, err := whisper.New(ModelPath)
+	assert.NoError(err)
+	assert.NotNil(model)
+	defer model.Close()
+
+	// Get context for decoding
+	ctx, err := model.NewContext()
+	assert.NoError(err)
+	assert.NotNil(ctx)
+
+}

bindings/go/pkg/whisper/doc.go

@@ -0,0 +1,4 @@
+/*
+This is the higher-level speech-to-text whisper.cpp API for go
+*/
+package whisper

bindings/go/pkg/whisper/interface.go

@@ -0,0 +1,93 @@
+package whisper
+
+import (
+	"io"
+	"time"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+// SegmentCallback is the callback function for processing segments in real
+// time. It is called during the Process function
+type SegmentCallback func(Segment)
+
+// Model is the interface to a whisper model. Create a new model with the
+// function whisper.New(string)
+type Model interface {
+	io.Closer
+
+	// Return a new speech-to-text context.
+	NewContext() (Context, error)
+
+	// Return true if the model is multilingual.
+	IsMultilingual() bool
+
+	// Return all languages supported.
+	Languages() []string
+}
+
+// Context is the speach recognition context.
+type Context interface {
+	SetLanguage(string) error // Set the language to use for speech recognition, use "auto" for auto detect language.
+	SetTranslate(bool)        // Set translate flag
+	IsMultilingual() bool     // Return true if the model is multilingual.
+	Language() string         // Get language
+
+	SetOffset(time.Duration)      // Set offset
+	SetDuration(time.Duration)    // Set duration
+	SetThreads(uint)              // Set number of threads to use
+	SetSpeedup(bool)              // Set speedup 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
+
+	// After process is called, return segments until the end of the stream
+	// is reached, when io.EOF is returned.
+	NextSegment() (Segment, error)
+
+	IsBEG(Token) bool          // Test for "begin" token
+	IsSOT(Token) bool          // Test for "start of transcription" token
+	IsEOT(Token) bool          // Test for "end of transcription" token
+	IsPREV(Token) bool         // Test for "start of prev" token
+	IsSOLM(Token) bool         // Test for "start of lm" token
+	IsNOT(Token) bool          // Test for "No timestamps" token
+	IsLANG(Token, string) bool // Test for token associated with a specific language
+	IsText(Token) bool         // Test for text token
+
+	// Timings
+	PrintTimings()
+	ResetTimings()
+
+	SystemInfo() string
+}
+
+// Segment is the text result of a speech recognition.
+type Segment struct {
+	// Segment Number
+	Num int
+
+	// Time beginning and end timestamps for the segment.
+	Start, End time.Duration
+
+	// The text of the segment.
+	Text string
+
+	// The tokens of the segment.
+	Tokens []Token
+}
+
+// Token is a text or special token
+type Token struct {
+	Id         int
+	Text       string
+	P          float32
+	Start, End time.Duration
+}

bindings/go/pkg/whisper/model.go

@@ -0,0 +1,101 @@
+package whisper
+
+import (
+	"fmt"
+	"os"
+	"runtime"
+
+	// Bindings
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type model struct {
+	path string
+	ctx  *whisper.Context
+}
+
+// Make sure model adheres to the interface
+var _ Model = (*model)(nil)
+
+///////////////////////////////////////////////////////////////////////////////
+// LIFECYCLE
+
+func New(path string) (Model, error) {
+	model := new(model)
+	if _, err := os.Stat(path); err != nil {
+		return nil, err
+	} else if ctx := whisper.Whisper_init(path); ctx == nil {
+		return nil, ErrUnableToLoadModel
+	} else {
+		model.ctx = ctx
+		model.path = path
+	}
+
+	// Return success
+	return model, nil
+}
+
+func (model *model) Close() error {
+	if model.ctx != nil {
+		model.ctx.Whisper_free()
+	}
+
+	// Release resources
+	model.ctx = nil
+
+	// Return success
+	return nil
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// STRINGIFY
+
+func (model *model) String() string {
+	str := "<whisper.model"
+	if model.ctx != nil {
+		str += fmt.Sprintf(" model=%q", model.path)
+	}
+	return str + ">"
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Return true if model is multilingual (language and translation options are supported)
+func (model *model) IsMultilingual() bool {
+	return model.ctx.Whisper_is_multilingual() != 0
+}
+
+// Return all recognized languages. Initially it is set to auto-detect
+func (model *model) Languages() []string {
+	result := make([]string, 0, whisper.Whisper_lang_max_id())
+	for i := 0; i < whisper.Whisper_lang_max_id(); i++ {
+		str := whisper.Whisper_lang_str(i)
+		if model.ctx.Whisper_lang_id(str) >= 0 {
+			result = append(result, str)
+		}
+	}
+	return result
+}
+
+func (model *model) NewContext() (Context, error) {
+	if model.ctx == nil {
+		return nil, ErrInternalAppError
+	}
+
+	// Create new context
+	params := model.ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY)
+	params.SetTranslate(false)
+	params.SetPrintSpecial(false)
+	params.SetPrintProgress(false)
+	params.SetPrintRealtime(false)
+	params.SetPrintTimestamps(false)
+	params.SetThreads(runtime.NumCPU())
+	params.SetNoContext(true)
+
+	// Return new context
+	return newContext(model, params)
+}

bindings/go/whisper.go

@@ -0,0 +1,417 @@
+package whisper
+
+import (
+	"errors"
+	"unsafe"
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// CGO
+
+/*
+#cgo LDFLAGS: -lwhisper -lm -lstdc++
+#cgo darwin LDFLAGS: -framework Accelerate
+#include <whisper.h>
+#include <stdlib.h>
+
+extern void callNewSegment(void* user_data, int new);
+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, struct whisper_state* state, int n_new, void* user_data) {
+    if(user_data != NULL && ctx != NULL) {
+        callNewSegment(user_data, n_new);
+    }
+}
+
+// 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, struct whisper_state* state, void* user_data) {
+    if(user_data != NULL && ctx != NULL) {
+        return callEncoderBegin(user_data);
+    }
+    return false;
+}
+
+// Get default parameters and set callbacks
+static struct whisper_full_params whisper_full_default_params_cb(struct whisper_context* ctx, enum whisper_sampling_strategy strategy) {
+	struct whisper_full_params params = whisper_full_default_params(strategy);
+	params.new_segment_callback = whisper_new_segment_cb;
+	params.new_segment_callback_user_data = (void*)(ctx);
+	params.encoder_begin_callback = whisper_encoder_begin_cb;
+	params.encoder_begin_callback_user_data = (void*)(ctx);
+	return params;
+}
+*/
+import "C"
+
+///////////////////////////////////////////////////////////////////////////////
+// TYPES
+
+type (
+	Context          C.struct_whisper_context
+	Token            C.whisper_token
+	TokenData        C.struct_whisper_token_data
+	SamplingStrategy C.enum_whisper_sampling_strategy
+	Params           C.struct_whisper_full_params
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// GLOBALS
+
+const (
+	SAMPLING_GREEDY      SamplingStrategy = C.WHISPER_SAMPLING_GREEDY
+	SAMPLING_BEAM_SEARCH SamplingStrategy = C.WHISPER_SAMPLING_BEAM_SEARCH
+)
+
+const (
+	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
+	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
+	NumFFT     = C.WHISPER_N_FFT
+	NumMEL     = C.WHISPER_N_MEL
+	HopLength  = C.WHISPER_HOP_LENGTH
+	ChunkSize  = C.WHISPER_CHUNK_SIZE
+)
+
+var (
+	ErrTokenizerFailed  = errors.New("whisper_tokenize failed")
+	ErrAutoDetectFailed = errors.New("whisper_lang_auto_detect failed")
+	ErrConversionFailed = errors.New("whisper_convert failed")
+	ErrInvalidLanguage  = errors.New("invalid language")
+)
+
+///////////////////////////////////////////////////////////////////////////////
+// PUBLIC METHODS
+
+// Allocates all memory needed for the model and loads the model from the given file.
+// Returns NULL on failure.
+func Whisper_init(path string) *Context {
+	cPath := C.CString(path)
+	defer C.free(unsafe.Pointer(cPath))
+	if ctx := C.whisper_init_from_file(cPath); ctx != nil {
+		return (*Context)(ctx)
+	} else {
+		return nil
+	}
+}
+
+// Frees all memory allocated by the model.
+func (ctx *Context) Whisper_free() {
+	C.whisper_free((*C.struct_whisper_context)(ctx))
+}
+
+// Convert RAW PCM audio to log mel spectrogram.
+// The resulting spectrogram is stored inside the provided whisper context.
+func (ctx *Context) Whisper_pcm_to_mel(data []float32, threads int) error {
+	if C.whisper_pcm_to_mel((*C.struct_whisper_context)(ctx), (*C.float)(&data[0]), C.int(len(data)), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// This can be used to set a custom log mel spectrogram inside 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
+func (ctx *Context) Whisper_set_mel(data []float32, n_mel int) error {
+	if C.whisper_set_mel((*C.struct_whisper_context)(ctx), (*C.float)(&data[0]), C.int(len(data)), C.int(n_mel)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// Run the Whisper encoder on the log mel spectrogram stored inside 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.
+func (ctx *Context) Whisper_encode(offset, threads int) error {
+	if C.whisper_encode((*C.struct_whisper_context)(ctx), C.int(offset), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// 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.
+func (ctx *Context) Whisper_decode(tokens []Token, past, threads int) error {
+	if C.whisper_decode((*C.struct_whisper_context)(ctx), (*C.whisper_token)(&tokens[0]), C.int(len(tokens)), C.int(past), C.int(threads)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// 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
+func (ctx *Context) Whisper_tokenize(text string, tokens []Token) (int, error) {
+	cText := C.CString(text)
+	defer C.free(unsafe.Pointer(cText))
+	if n := C.whisper_tokenize((*C.struct_whisper_context)(ctx), cText, (*C.whisper_token)(&tokens[0]), C.int(len(tokens))); n >= 0 {
+		return int(n), nil
+	} else {
+		return 0, ErrTokenizerFailed
+	}
+}
+
+// Return the id of the specified language, returns -1 if not found
+// Examples:
+//
+//	"de" -> 2
+//	"german" -> 2
+func (ctx *Context) Whisper_lang_id(lang string) int {
+	return int(C.whisper_lang_id(C.CString(lang)))
+}
+
+// Largest language id (i.e. number of available languages - 1)
+func Whisper_lang_max_id() int {
+	return int(C.whisper_lang_max_id())
+}
+
+// Return the short string of the specified language id (e.g. 2 -> "de"),
+// returns empty string if not found
+func Whisper_lang_str(id int) string {
+	return C.GoString(C.whisper_lang_str(C.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 probabilities of all languages.
+// ref: https://github.com/openai/whisper/blob/main/whisper/decoding.py#L18-L69
+func (ctx *Context) Whisper_lang_auto_detect(offset_ms, n_threads int) ([]float32, error) {
+	probs := make([]float32, Whisper_lang_max_id()+1)
+	if n := int(C.whisper_lang_auto_detect((*C.struct_whisper_context)(ctx), C.int(offset_ms), C.int(n_threads), (*C.float)(&probs[0]))); n < 0 {
+		return nil, ErrAutoDetectFailed
+	} else {
+		return probs, nil
+	}
+}
+
+func (ctx *Context) Whisper_n_len() int {
+	return int(C.whisper_n_len((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_n_vocab() int {
+	return int(C.whisper_n_vocab((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_n_text_ctx() int {
+	return int(C.whisper_n_text_ctx((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_n_audio_ctx() int {
+	return int(C.whisper_n_audio_ctx((*C.struct_whisper_context)(ctx)))
+}
+
+func (ctx *Context) Whisper_is_multilingual() int {
+	return int(C.whisper_is_multilingual((*C.struct_whisper_context)(ctx)))
+}
+
+// The probabilities for the next token
+//func (ctx *Whisper_context) Whisper_get_probs() []float32 {
+//	return (*[1 << 30]float32)(unsafe.Pointer(C.whisper_get_probs((*C.struct_whisper_context)(ctx))))[:ctx.Whisper_n_vocab()]
+//}
+
+// Token Id -> String. Uses the vocabulary in the provided context
+func (ctx *Context) Whisper_token_to_str(token Token) string {
+	return C.GoString(C.whisper_token_to_str((*C.struct_whisper_context)(ctx), C.whisper_token(token)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_eot() Token {
+	return Token(C.whisper_token_eot((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_sot() Token {
+	return Token(C.whisper_token_sot((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_prev() Token {
+	return Token(C.whisper_token_prev((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_solm() Token {
+	return Token(C.whisper_token_solm((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_not() Token {
+	return Token(C.whisper_token_not((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_beg() Token {
+	return Token(C.whisper_token_beg((*C.struct_whisper_context)(ctx)))
+}
+
+// Special tokens
+func (ctx *Context) Whisper_token_lang(lang_id int) Token {
+	return Token(C.whisper_token_lang((*C.struct_whisper_context)(ctx), C.int(lang_id)))
+}
+
+// Task tokens
+func Whisper_token_translate() Token {
+	return Token(C.whisper_token_translate())
+}
+
+// Task tokens
+func Whisper_token_transcribe() Token {
+	return Token(C.whisper_token_transcribe())
+}
+
+// Performance information
+func (ctx *Context) Whisper_print_timings() {
+	C.whisper_print_timings((*C.struct_whisper_context)(ctx))
+}
+
+// Performance information
+func (ctx *Context) Whisper_reset_timings() {
+	C.whisper_reset_timings((*C.struct_whisper_context)(ctx))
+}
+
+// Print system information
+func Whisper_print_system_info() string {
+	return C.GoString(C.whisper_print_system_info())
+}
+
+// Return default parameters for a strategy
+func (ctx *Context) Whisper_full_default_params(strategy SamplingStrategy) Params {
+	// Get default parameters
+	return Params(C.whisper_full_default_params_cb((*C.struct_whisper_context)(ctx), C.enum_whisper_sampling_strategy(strategy)))
+}
+
+// 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 {
+	registerEncoderBeginCallback(ctx, encoderBeginCallback)
+	registerNewSegmentCallback(ctx, newSegmentCallback)
+	defer registerEncoderBeginCallback(ctx, nil)
+	defer registerNewSegmentCallback(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 {
+		return ErrConversionFailed
+	}
+}
+
+// Split the input audio in chunks and process each chunk separately using whisper_full()
+// 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.
+func (ctx *Context) Whisper_full_parallel(params Params, samples []float32, processors int, encoderBeginCallback func() bool, newSegmentCallback func(int)) error {
+	registerEncoderBeginCallback(ctx, encoderBeginCallback)
+	registerNewSegmentCallback(ctx, newSegmentCallback)
+	defer registerEncoderBeginCallback(ctx, nil)
+	defer registerNewSegmentCallback(ctx, nil)
+
+	if C.whisper_full_parallel((*C.struct_whisper_context)(ctx), (C.struct_whisper_full_params)(params), (*C.float)(&samples[0]), C.int(len(samples)), C.int(processors)) == 0 {
+		return nil
+	} else {
+		return ErrConversionFailed
+	}
+}
+
+// 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 {
+	return int(C.whisper_full_n_segments((*C.struct_whisper_context)(ctx)))
+}
+
+// Get the start and end time of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_t0(segment int) int64 {
+	return int64(C.whisper_full_get_segment_t0((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the start and end time of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_t1(segment int) int64 {
+	return int64(C.whisper_full_get_segment_t1((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the text of the specified segment.
+func (ctx *Context) Whisper_full_get_segment_text(segment int) string {
+	return C.GoString(C.whisper_full_get_segment_text((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get number of tokens in the specified segment.
+func (ctx *Context) Whisper_full_n_tokens(segment int) int {
+	return int(C.whisper_full_n_tokens((*C.struct_whisper_context)(ctx), C.int(segment)))
+}
+
+// Get the token text of the specified token index in the specified segment.
+func (ctx *Context) Whisper_full_get_token_text(segment int, token int) string {
+	return C.GoString(C.whisper_full_get_token_text((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+// Get the token of the specified token index in the specified segment.
+func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {
+	return Token(C.whisper_full_get_token_id((*C.struct_whisper_context)(ctx), C.int(segment), C.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 {
+	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+// Get the probability of the specified token in the specified segment.
+func (ctx *Context) Whisper_full_get_token_p(segment int, token int) float32 {
+	return float32(C.whisper_full_get_token_p((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// CALLBACKS
+
+var (
+	cbNewSegment   = make(map[unsafe.Pointer]func(int))
+	cbEncoderBegin = make(map[unsafe.Pointer]func() bool)
+)
+
+func registerNewSegmentCallback(ctx *Context, fn func(int)) {
+	if fn == nil {
+		delete(cbNewSegment, unsafe.Pointer(ctx))
+	} else {
+		cbNewSegment[unsafe.Pointer(ctx)] = fn
+	}
+}
+
+func registerEncoderBeginCallback(ctx *Context, fn func() bool) {
+	if fn == nil {
+		delete(cbEncoderBegin, unsafe.Pointer(ctx))
+	} else {
+		cbEncoderBegin[unsafe.Pointer(ctx)] = fn
+	}
+}
+
+//export callNewSegment
+func callNewSegment(user_data unsafe.Pointer, new C.int) {
+	if fn, ok := cbNewSegment[user_data]; ok {
+		fn(int(new))
+	}
+}
+
+//export callEncoderBegin
+func callEncoderBegin(user_data unsafe.Pointer) C.bool {
+	if fn, ok := cbEncoderBegin[user_data]; ok {
+		if fn() {
+			return C.bool(true)
+		} else {
+			return C.bool(false)
+		}
+	}
+	return true
+}
+
+func (t TokenData) T0() int64 {
+	return int64(t.t0)
+}
+
+func (t TokenData) T1() int64 {
+	return int64(t.t1)
+}

bindings/go/whisper_test.go

@@ -0,0 +1,113 @@
+package whisper_test
+
+import (
+	"os"
+	"runtime"
+	"testing"
+	"time"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go"
+	wav "github.com/go-audio/wav"
+	assert "github.com/stretchr/testify/assert"
+)
+
+const (
+	ModelPath  = "models/ggml-small.en.bin"
+	SamplePath = "samples/jfk.wav"
+)
+
+func Test_Whisper_000(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	ctx.Whisper_free()
+}
+
+func Test_Whisper_001(t *testing.T) {
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Open samples
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Read samples
+	d := wav.NewDecoder(fh)
+	buf, err := d.FullPCMBuffer()
+	assert.NoError(err)
+
+	// Run whisper
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	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)
+	assert.NoError(err)
+
+	// Print out tokens
+	num_segments := ctx.Whisper_full_n_segments()
+	assert.GreaterOrEqual(num_segments, 1)
+	for i := 0; i < num_segments; i++ {
+		str := ctx.Whisper_full_get_segment_text(i)
+		assert.NotEmpty(str)
+		t0 := time.Duration(ctx.Whisper_full_get_segment_t0(i)) * time.Millisecond
+		t1 := time.Duration(ctx.Whisper_full_get_segment_t1(i)) * time.Millisecond
+		t.Logf("[%6s->%-6s] %q", t0, t1, str)
+	}
+}
+
+func Test_Whisper_002(t *testing.T) {
+	assert := assert.New(t)
+	for i := 0; i < whisper.Whisper_lang_max_id(); i++ {
+		str := whisper.Whisper_lang_str(i)
+		assert.NotEmpty(str)
+		t.Log(str)
+	}
+}
+
+func Test_Whisper_003(t *testing.T) {
+	threads := runtime.NumCPU()
+	assert := assert.New(t)
+	if _, err := os.Stat(ModelPath); os.IsNotExist(err) {
+		t.Skip("Skipping test, model not found:", ModelPath)
+	}
+	if _, err := os.Stat(SamplePath); os.IsNotExist(err) {
+		t.Skip("Skipping test, sample not found:", SamplePath)
+	}
+
+	// Open samples
+	fh, err := os.Open(SamplePath)
+	assert.NoError(err)
+	defer fh.Close()
+
+	// Read samples
+	d := wav.NewDecoder(fh)
+	buf, err := d.FullPCMBuffer()
+	assert.NoError(err)
+
+	// Make the model
+	ctx := whisper.Whisper_init(ModelPath)
+	assert.NotNil(ctx)
+	defer ctx.Whisper_free()
+
+	// Get MEL
+	assert.NoError(ctx.Whisper_pcm_to_mel(buf.AsFloat32Buffer().Data, threads))
+
+	// Get Languages
+	languages, err := ctx.Whisper_lang_auto_detect(0, threads)
+	assert.NoError(err)
+	for i, p := range languages {
+		t.Logf("%s: %f", whisper.Whisper_lang_str(i), p)
+	}
+}

bindings/javascript/CMakeLists.txt

@@ -9,25 +9,33 @@ target_link_libraries(${TARGET} PRIVATE
     )
 
 unset(EXTRA_FLAGS)
+
 if (WHISPER_WASM_SINGLE_FILE)
     set(EXTRA_FLAGS "-s SINGLE_FILE=1")
     message(STATUS "Embedding WASM inside whisper.js")
 
     add_custom_command(
-        TARGET libwhisper POST_BUILD
+        TARGET ${TARGET} POST_BUILD
         COMMAND ${CMAKE_COMMAND} -E copy
         ${CMAKE_BINARY_DIR}/bin/libwhisper.js
         ${CMAKE_CURRENT_SOURCE_DIR}/whisper.js
         )
+
+    add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_BINARY_DIR}/bin/libwhisper.worker.js
+        ${CMAKE_CURRENT_SOURCE_DIR}/libwhisper.worker.js
+        )
 endif()
 
 set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
+    -s MODULARIZE=1 \
+    -s EXPORT_NAME=\"'whisper_factory'\" \
+    -s FORCE_FILESYSTEM=1 \
     -s USE_PTHREADS=1 \
     -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1610612736 \
-    -s TOTAL_MEMORY=1610612736 \
-    -s FORCE_FILESYSTEM=1 \
-    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    -s ALLOW_MEMORY_GROWTH=1 \
     ${EXTRA_FLAGS} \
     ")

bindings/javascript/README.md

@@ -0,0 +1,78 @@
+# whisper.cpp
+
+Node.js package for Whisper speech recognition
+
+Package: https://www.npmjs.com/package/whisper.cpp
+
+## Details
+
+The performance is comparable to when running `whisper.cpp` in the browser via WASM.
+
+The API is currently very rudimentary: [bindings/javascript/emscripten.cpp](/bindings/javascript/emscripten.cpp)
+
+For sample usage check [tests/test-whisper.js](/tests/test-whisper.js)
+
+## Package building + test
+
+```bash
+# load emscripten
+source /path/to/emsdk/emsdk_env.sh
+
+# clone repo
+git clone https://github.com/ggerganov/whisper.cpp
+cd whisper.cpp
+
+# grab base.en model
+./models/download-ggml-model.sh base.en
+
+# prepare PCM sample for testing
+ffmpeg -i samples/jfk.wav -f f32le -acodec pcm_f32le samples/jfk.pcmf32
+
+# build
+mkdir build-em && cd build-em
+emcmake cmake .. && make -j
+
+# run test
+node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js
+
+# publish npm package
+make publish-npm
+```
+
+## Sample run
+
+```java
+$ node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js
+
+whisper_model_load: loading model from 'whisper.bin'
+whisper_model_load: n_vocab       = 51864
+whisper_model_load: n_audio_ctx   = 1500
+whisper_model_load: n_audio_state = 512
+whisper_model_load: n_audio_head  = 8
+whisper_model_load: n_audio_layer = 6
+whisper_model_load: n_text_ctx    = 448
+whisper_model_load: n_text_state  = 512
+whisper_model_load: n_text_head   = 8
+whisper_model_load: n_text_layer  = 6
+whisper_model_load: n_mels        = 80
+whisper_model_load: f16           = 1
+whisper_model_load: type          = 2
+whisper_model_load: adding 1607 extra tokens
+whisper_model_load: mem_required  =  506.00 MB
+whisper_model_load: ggml ctx size =  140.60 MB
+whisper_model_load: memory size   =   22.83 MB
+whisper_model_load: model size    =  140.54 MB
+
+system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 | 
+
+operator(): processing 176000 samples, 11.0 sec, 8 threads, 1 processors, lang = en, task = transcribe ...
+
+[00:00:00.000 --> 00:00:11.000]   And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
+
+whisper_print_timings:     load time =   162.37 ms
+whisper_print_timings:      mel time =   183.70 ms
+whisper_print_timings:   sample time =     4.27 ms
+whisper_print_timings:   encode time =  8582.63 ms / 1430.44 ms per layer
+whisper_print_timings:   decode time =   436.16 ms / 72.69 ms per layer
+whisper_print_timings:    total time =  9370.90 ms
+```

bindings/javascript/emscripten.cpp

@@ -1,59 +1,58 @@
+//
+// This is the Javascript API of whisper.cpp
+//
+// Very crude at the moment.
+// Feel free to contribute and make this better!
+//
+// See the tests/test-whisper.js for sample usage
+//
+
 #include "whisper.h"
 
 #include <emscripten.h>
 #include <emscripten/bind.h>
 
-#include <vector>
 #include <thread>
+#include <vector>
 
-std::vector<struct whisper_context *> g_contexts(4, nullptr);
+struct whisper_context * g_context;
 
 EMSCRIPTEN_BINDINGS(whisper) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
-        for (size_t i = 0; i < g_contexts.size(); ++i) {
-            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
-                if (g_contexts[i] != nullptr) {
-                    return i + 1;
-                } else {
-                    return (size_t) 0;
-                }
+        if (g_context == nullptr) {
+            g_context = whisper_init_from_file(path_model.c_str());
+            if (g_context != nullptr) {
+                return true;
+            } else {
+                return false;
             }
         }
 
-        return (size_t) 0;
+        return false;
     }));
 
-    emscripten::function("free", emscripten::optional_override([](size_t index) {
-        --index;
-
-        if (index < g_contexts.size()) {
-            whisper_free(g_contexts[index]);
-            g_contexts[index] = nullptr;
+    emscripten::function("free", emscripten::optional_override([]() {
+        if (g_context) {
+            whisper_free(g_context);
+            g_context = nullptr;
         }
     }));
 
-    emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, bool translate) {
-        --index;
-
-        if (index >= g_contexts.size()) {
+    emscripten::function("full_default", emscripten::optional_override([](const emscripten::val & audio, const std::string & lang, bool translate) {
+        if (g_context == nullptr) {
             return -1;
         }
 
-        if (g_contexts[index] == nullptr) {
-            return -2;
-        }
-
         struct whisper_full_params params = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
 
-        params.print_realtime       = true;
-        params.print_progress       = false;
-        params.print_timestamps     = true;
-        params.print_special_tokens = false;
-        params.translate            = translate;
-        params.language             = whisper_is_multilingual(g_contexts[index]) ? lang.c_str() : "en";
-        params.n_threads            = std::min(8, (int) std::thread::hardware_concurrency());
-        params.offset_ms            = 0;
+        params.print_realtime   = true;
+        params.print_progress   = false;
+        params.print_timestamps = true;
+        params.print_special    = false;
+        params.translate        = translate;
+        params.language         = whisper_is_multilingual(g_context) ? lang.c_str() : "en";
+        params.n_threads        = std::min(8, (int) std::thread::hardware_concurrency());
+        params.offset_ms        = 0;
 
         std::vector<float> pcmf32;
         const int n = audio["length"].as<int>();
@@ -68,9 +67,11 @@ EMSCRIPTEN_BINDINGS(whisper) {
 
         // print system information
         {
+            printf("\n");
             printf("system_info: n_threads = %d / %d | %s\n",
                     params.n_threads, std::thread::hardware_concurrency(), whisper_print_system_info());
 
+            printf("\n");
             printf("%s: processing %d samples, %.1f sec, %d threads, %d processors, lang = %s, task = %s ...\n",
                     __func__, int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                     params.n_threads, 1,
@@ -80,10 +81,13 @@ EMSCRIPTEN_BINDINGS(whisper) {
             printf("\n");
         }
 
-        int ret = whisper_full(g_contexts[index], params, pcmf32.data(), pcmf32.size());
+        // run whisper
+        {
+            whisper_reset_timings(g_context);
+            whisper_full(g_context, params, pcmf32.data(), pcmf32.size());
+            whisper_print_timings(g_context);
+        }
 
-        whisper_print_timings(g_contexts[index]);
-
-        return ret;
+        return 0;
     }));
 }

bindings/javascript/libwhisper.worker.js

@@ -0,0 +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}};

bindings/javascript/package-tmpl.json

@@ -0,0 +1,26 @@
+{
+  "name": "whisper.cpp",
+  "version": "@PROJECT_VERSION@",
+  "description": "Whisper speech recognition",
+  "main": "whisper.js",
+  "scripts": {
+    "test": "echo \"todo: add tests\" && exit 0"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/ggerganov/whisper.cpp"
+  },
+  "keywords": [
+    "openai",
+    "whisper",
+    "speech-to-text",
+    "speech-recognition",
+    "transformer"
+  ],
+  "author": "Georgi Gerganov",
+  "license": "MIT",
+  "bugs": {
+    "url": "https://github.com/ggerganov/whisper.cpp/issues"
+  },
+  "homepage": "https://github.com/ggerganov/whisper.cpp#readme"
+}

bindings/javascript/package.json

@@ -0,0 +1,26 @@
+{
+  "name": "whisper.cpp",
+  "version": "1.3.0",
+  "description": "Whisper speech recognition",
+  "main": "whisper.js",
+  "scripts": {
+    "test": "echo \"todo: add tests\" && exit 0"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/ggerganov/whisper.cpp"
+  },
+  "keywords": [
+    "openai",
+    "whisper",
+    "speech-to-text",
+    "speech-recognition",
+    "transformer"
+  ],
+  "author": "Georgi Gerganov",
+  "license": "MIT",
+  "bugs": {
+    "url": "https://github.com/ggerganov/whisper.cpp/issues"
+  },
+  "homepage": "https://github.com/ggerganov/whisper.cpp#readme"
+}

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

cmake/DefaultTargetOptions.cmake

@@ -0,0 +1,17 @@
+# Set the default compile features and properties for a target.
+
+if (NOT TARGET)
+    message(FATAL_ERROR "TARGET not set before including DefaultTargetOptions")
+endif()
+
+target_compile_features(${TARGET}
+    PRIVATE
+        cxx_std_11
+    )
+
+set_target_properties(${TARGET}
+    PROPERTIES
+        EXPORT_COMPILE_COMMANDS ON
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin"
+        INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib"
+)

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 +1,67 @@
+#import "coreml/whisper-encoder.h"
+#import "coreml/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];
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model 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
+    ];
+
+    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+
+    MLMultiArray * outMA = outCoreML.output;
+
+    //NSArray<NSNumber *> * shape = outMA.shape;
+    //NSArray<NSNumber *> * strides = outMA.strides;
+
+    //printf("shape:   %ld %ld %ld %ld\n", [shape[0] longValue], [shape[1] longValue], [shape[2] longValue], [shape[3] longValue]);
+    //printf("strides: %ld %ld %ld %ld\n", [strides[0] longValue], [strides[1] longValue], [strides[2] longValue], [strides[3] longValue]);
+
+    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
+}
+
+#if __cplusplus
+}
+#endif

examples/CMakeLists.txt

@@ -14,14 +14,54 @@ if (WHISPER_SUPPORT_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
+# common
+
+set(TARGET common)
+
+add_library(${TARGET} STATIC
+    common.h
+    common.cpp
+    )
+
+include(DefaultTargetOptions)
+
+set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+if (WHISPER_SUPPORT_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})
 
 if (EMSCRIPTEN)
     add_subdirectory(whisper.wasm)
+    add_subdirectory(stream.wasm)
+    add_subdirectory(command.wasm)
+    add_subdirectory(talk.wasm)
+    add_subdirectory(bench.wasm)
+elseif(CMAKE_JS_VERSION)
+    add_subdirectory(addon.node)
 else()
     add_subdirectory(main)
     add_subdirectory(stream)
+    add_subdirectory(command)
     add_subdirectory(bench)
+    add_subdirectory(talk)
+    add_subdirectory(talk-llama)
 endif()

examples/addon.node/.gitignore

@@ -0,0 +1,3 @@
+.idea
+node_modules
+build

examples/addon.node/CMakeLists.txt

@@ -0,0 +1,31 @@
+set(TARGET whisper-addon)
+
+# Base settings
+#==================================================================
+# env var supported by cmake-js
+add_definitions(-DNAPI_VERSION=4)
+include_directories(${CMAKE_JS_INC})
+#==================================================================
+
+add_library(${TARGET} SHARED ${CMAKE_JS_SRC} addon.cpp)
+set_target_properties(${TARGET} PROPERTIES PREFIX "" SUFFIX ".node")
+
+include(DefaultTargetOptions)
+
+# Include N-API wrappers
+#==================================================================
+execute_process(COMMAND node -p "require('node-addon-api').include"
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE NODE_ADDON_API_DIR
+        )
+string(REPLACE "\n" "" NODE_ADDON_API_DIR ${NODE_ADDON_API_DIR})
+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} 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

@@ -0,0 +1,37 @@
+# addon
+
+This is an addon demo that can **perform whisper model reasoning in `node` and `electron` environments**, based on [cmake-js](https://github.com/cmake-js/cmake-js).
+It can be used as a reference for using the whisper.cpp project in other node projects.
+
+## Install
+
+```shell
+npm install
+```
+
+## Compile
+
+Make sure it is in the project root directory and compiled with make-js.
+
+```shell
+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 -B Release
+> ```
+
+## Run
+
+```shell
+cd examples/addon.node
+
+node index.js --language='language' --model='model-path' --fname_inp='file-path'
+```
+
+Because this is a simple Demo, only the above parameters are set in the node environment.
+
+Other parameters can also be specified in the node environment.

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

@@ -0,0 +1,22 @@
+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);
+  });
+});

examples/addon.node/addon.cpp

@@ -0,0 +1,338 @@
+#include "napi.h"
+#include "common.h"
+
+#include "whisper.h"
+
+#include <string>
+#include <thread>
+#include <vector>
+#include <cmath>
+#include <cstdint>
+
+struct whisper_params {
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors = 1;
+    int32_t offset_t_ms  = 0;
+    int32_t offset_n     = 0;
+    int32_t duration_ms  = 0;
+    int32_t max_context  = -1;
+    int32_t max_len      = 0;
+    int32_t best_of      = 5;
+    int32_t beam_size    = -1;
+
+    float word_thold    = 0.01f;
+    float entropy_thold = 2.4f;
+    float logprob_thold = -1.0f;
+
+    bool speed_up       = false;
+    bool translate      = false;
+    bool diarize        = 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;
+
+    std::string language = "en";
+    std::string prompt;
+    std::string model    = "../../ggml-large.bin";
+
+    std::vector<std::string> fname_inp = {};
+    std::vector<std::string> fname_out = {};
+};
+
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+};
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+
+    std::string speaker = "";
+
+    int64_t t0;
+    int64_t t1;
+
+    // print the last n_new segments
+    const int s0 = n_segments - n_new;
+
+    if (s0 == 0) {
+        printf("\n");
+    }
+
+    for (int i = s0; i < n_segments; i++) {
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
+        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            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());
+        }
+
+        // colorful print bug
+        //
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        printf("%s%s", speaker.c_str(), text);
+
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+
+        fflush(stdout);
+    }
+}
+
+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;
+    }
+
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 3;
+    }
+
+    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
+        const auto fname_inp = params.fname_inp[f];
+        const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
+
+        std::vector<float> pcmf32; // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+        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
+        {
+            fprintf(stderr, "\n");
+            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
+        }
+
+        // print some info about the processing
+        {
+            fprintf(stderr, "\n");
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %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.no_timestamps ? 0 : 1);
+
+            fprintf(stderr, "\n");
+        }
+
+        // run the inference
+        {
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+
+            wparams.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.token_timestamps = params.output_wts || params.max_len > 0;
+            wparams.thold_pt         = params.word_thold;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+
+            wparams.speed_up         = params.speed_up;
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.initial_prompt   = params.prompt.c_str();
+
+            whisper_print_user_data user_data = { &params, &pcmf32s };
+
+            // this callback is called on each new segment
+            if (!wparams.print_realtime) {
+                wparams.new_segment_callback           = whisper_print_segment_callback;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            // 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*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
+            }
+
+            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                fprintf(stderr, "failed to process audio\n");
+                return 10;
+            }
+        }
+    }
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    result.resize(n_segments);
+    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);
+
+        result[i].emplace_back(to_timestamp(t0, true));
+        result[i].emplace_back(to_timestamp(t1, true));
+        result[i].emplace_back(text);
+    }
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}
+
+class Worker : public Napi::AsyncWorker {
+ public:
+  Worker(Napi::Function& callback, whisper_params params)
+      : Napi::AsyncWorker(callback), params(params) {}
+
+  void Execute() override {
+    run(params, result);
+  }
+
+  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});
+  }
+
+ private:
+  whisper_params params;
+  std::vector<std::vector<std::string>> result;
+};
+
+
+
+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;
+}
+
+NODE_API_MODULE(whisper, Init);

examples/addon.node/index.js

@@ -0,0 +1,36 @@
+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: "../../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;
+  }, [])
+);
+
+for (const key in params) {
+  if (whisperParams.hasOwnProperty(key)) {
+    whisperParams[key] = params[key];
+  }
+}
+
+console.log("whisperParams =", whisperParams);
+
+whisperAsync(whisperParams).then((result) => {
+  console.log(`Result from whisper: ${result}`);
+});

examples/addon.node/package.json

@@ -0,0 +1,16 @@
+{
+  "name": "whisper-addon",
+  "version": "0.0.0",
+  "description": "",
+  "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

@@ -0,0 +1,49 @@
+#
+# libbench
+#
+
+set(TARGET libbench)
+
+add_executable(${TARGET}
+    emscripten.cpp
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE
+    whisper
+    )
+
+unset(EXTRA_FLAGS)
+
+if (WHISPER_WASM_SINGLE_FILE)
+    set(EXTRA_FLAGS "-s SINGLE_FILE=1")
+    message(STATUS "Embedding WASM inside bench.js")
+
+    add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_BINARY_DIR}/bin/libbench.js
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/bench.wasm/bench.js
+        )
+endif()
+
+set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
+    --bind \
+    -s USE_PTHREADS=1 \
+    -s PTHREAD_POOL_SIZE=8 \
+    -s INITIAL_MEMORY=1024MB \
+    -s TOTAL_MEMORY=1024MB \
+    -s FORCE_FILESYSTEM=1 \
+    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    ${EXTRA_FLAGS} \
+    ")
+
+#
+# bench.wasm
+#
+
+set(TARGET bench.wasm)
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/../helpers.js    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/helpers.js @ONLY)

examples/bench.wasm/README.md

@@ -0,0 +1,22 @@
+# bench.wasm
+
+Benchmark the performance of whisper.cpp in the browser using WebAssembly
+
+Link: https://whisper.ggerganov.com/bench/
+
+Terminal version: [examples/bench](/examples/bench)
+
+## Build instructions
+
+```bash
+# build using Emscripten (v3.1.2)
+git clone https://github.com/ggerganov/whisper.cpp
+cd whisper.cpp
+mkdir build-em && cd build-em
+emcmake cmake ..
+make -j
+
+# copy the produced page to your HTTP path
+cp bin/bench.wasm/*       /path/to/html/
+cp bin/libbench.worker.js /path/to/html/
+```

examples/bench.wasm/emscripten.cpp

@@ -0,0 +1,85 @@
+#include "whisper.h"
+
+#include <emscripten.h>
+#include <emscripten/bind.h>
+
+#include <cmath>
+#include <string>
+#include <thread>
+#include <vector>
+
+constexpr int N_THREAD = 8;
+
+// TODO: get rid of this vector of contexts - bad idea in the first place
+std::vector<struct whisper_context *> g_contexts(4, nullptr);
+
+std::thread g_worker;
+
+void bench_main(size_t index) {
+    const int n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+
+    // whisper context
+    auto & ctx = g_contexts[index];
+
+    fprintf(stderr, "%s: running benchmark with %d threads - please wait...\n", __func__, n_threads);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
+        fprintf(stderr, "error: failed to set mel: %d\n", ret);
+        return;
+    }
+
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n", n_threads, std::thread::hardware_concurrency(), whisper_print_system_info());
+    }
+
+    if (int ret = whisper_encode(ctx, 0, n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        return;
+    }
+
+    whisper_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "If you wish, you can submit these results here:\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "  https://github.com/ggerganov/whisper.cpp/issues/89\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "Please include the following information:\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "  - CPU model\n");
+    fprintf(stderr, "  - Operating system\n");
+    fprintf(stderr, "  - Browser\n");
+    fprintf(stderr, "\n");
+}
+
+EMSCRIPTEN_BINDINGS(bench) {
+    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
+        for (size_t i = 0; i < g_contexts.size(); ++i) {
+            if (g_contexts[i] == nullptr) {
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                if (g_contexts[i] != nullptr) {
+                    if (g_worker.joinable()) {
+                        g_worker.join();
+                    }
+                    g_worker = std::thread([i]() {
+                        bench_main(i);
+                    });
+
+                    return i + 1;
+                } else {
+                    return (size_t) 0;
+                }
+            }
+        }
+
+        return (size_t) 0;
+    }));
+
+    emscripten::function("free", emscripten::optional_override([](size_t index) {
+        if (index < g_contexts.size()) {
+            whisper_free(g_contexts[index]);
+            g_contexts[index] = nullptr;
+        }
+    }));
+}

examples/bench.wasm/index-tmpl.html

@@ -0,0 +1,227 @@
+<!doctype html>
+<html lang="en-us">
+    <head>
+        <title>bench : Benchmark whisper.cpp performance in the browser</title>
+
+        <style>
+            #output {
+                width: 100%;
+                height: 100%;
+                margin: 0 auto;
+                margin-top: 10px;
+                border-left: 0px;
+                border-right: 0px;
+                padding-left: 0px;
+                padding-right: 0px;
+                display: block;
+                background-color: black;
+                color: white;
+                font-size: 10px;
+                font-family: 'Lucida Console', Monaco, monospace;
+                outline: none;
+                white-space: pre;
+                overflow-wrap: normal;
+                overflow-x: scroll;
+            }
+        </style>
+    </head>
+    <body>
+        <div id="main-container">
+            <b>bench : Benchmark whisper.cpp performance in the browser</b>
+
+            <br><br>
+
+            You can find more about this project on <a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/bench.wasm">GitHub</a>.
+
+            <br><br>
+
+            <hr>
+
+            Select the model you would like to use and click the "Bench" button.<br>
+            The results will be displayed in the textarea below.
+
+            <br><br>
+
+            <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>
+
+                <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+            </div>
+
+            <br>
+
+            <div id="input">
+                <button id="bench" onclick="onBench()" disabled>Bench</button>
+                <button id="clear" onclick="clearCache()">Clear Cache</button>
+            </div>
+
+            <hr>
+
+            Debug output:
+            <textarea id="output" rows="20"></textarea>
+
+            <br>
+
+            <b>Troubleshooting</b>
+
+            <br><br>
+
+            The page does some heavy computations, so make sure:
+
+            <ul>
+                <li>To use a modern web browser (e.g. Chrome, Firefox)</li>
+                <li>To use a fast desktop or laptop computer (i.e. not a mobile phone)</li>
+                <li>Your browser supports WASM <a href="https://webassembly.org/roadmap/">Fixed-width SIMD</a></li>
+            </ul>
+
+            <div class="cell-version">
+                <span>
+                    |
+                    Build time: <span class="nav-link">@GIT_DATE@</span> |
+                    Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
+                    Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
+                    <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/bench.wasm">Source Code</a> |
+                </span>
+            </div>
+        </div>
+
+        <script type="text/javascript" src="helpers.js"></script>
+        <script type='text/javascript'>
+            // the bench instance
+            var instance = null;
+
+            // model name
+            var model_whisper = null;
+
+            var Module = {
+                print: printTextarea,
+                printErr: printTextarea,
+                setStatus: function(text) {
+                    printTextarea('js: ' + text);
+                },
+                monitorRunDependencies: function(left) {
+                },
+                preRun: function() {
+                    printTextarea('js: Preparing ...');
+                },
+                postRun: function() {
+                    printTextarea('js: Initialized successfully!');
+                }
+            };
+
+            //
+            // fetch models
+            //
+
+            let dbVersion = 1
+            let dbName    = 'whisper.ggerganov.com';
+            let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
+
+            function storeFS(fname, buf) {
+                // write to WASM file using FS_createDataFile
+                // if the file exists, delete it
+                try {
+                    Module.FS_unlink(fname);
+                } catch (e) {
+                    // ignore
+                }
+
+                Module.FS_createDataFile("/", fname, buf, true, true);
+
+                printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
+
+                model_whisper = fname;
+
+                document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
+
+                if (model_whisper != null) {
+                    document.getElementById('bench').disabled = false;
+                }
+            }
+
+            function loadFile(event, fname) {
+                var file = event.target.files[0] || null;
+                if (file == null) {
+                    return;
+                }
+
+                printTextarea("loadFile: loading model: " + file.name + ", size: " + file.size + " bytes");
+                printTextarea('loadFile: please wait ...');
+
+                var reader = new FileReader();
+                reader.onload = function(event) {
+                    var buf = new Uint8Array(reader.result);
+                    storeFS(fname, buf);
+                }
+                reader.readAsArrayBuffer(file);
+
+                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('whisper-file'         ).style.display = 'none';
+                document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
+            }
+
+            function loadWhisper(model) {
+                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',
+                };
+
+                let sizes = {
+                    'tiny.en': 75,
+                    'base.en': 142,
+                };
+
+                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('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
+
+                cbProgress = function(p) {
+                    let el = document.getElementById('fetch-whisper-progress');
+                    el.innerHTML = Math.round(100*p) + '%';
+                };
+
+                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 = '';
+                };
+
+                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
+            }
+
+            //
+            // main
+            //
+
+            function onBench() {
+                if (instance) {
+                    Module.free(instance);
+                }
+
+                instance = Module.init('whisper.bin');
+
+                if (instance) {
+                    printTextarea("js: whisper initialized, instance: " + instance);
+                }
+
+                document.getElementById('bench').disabled = true;
+
+                if (!instance) {
+                    printTextarea("js: failed to initialize whisper");
+                    return;
+                }
+            }
+
+        </script>
+        <script type="text/javascript" src="bench.js"></script>
+    </body>
+</html>

examples/bench/CMakeLists.txt

@@ -1,3 +1,6 @@
 set(TARGET bench)
 add_executable(${TARGET} bench.cpp)
+
+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})

examples/bench/README.md

@@ -1,6 +1,8 @@
 # bench
 
-A very basic tool for benchmarking the inference performance on your device. The tool simply runs the Encoder part of the transformer on some random audio data and records the execution time. This way we can have an objective comparison of the performance of the model for various setups.
+A very basic tool for benchmarking the inference performance on your device. The tool simply runs the Encoder part of
+the transformer on some random audio data and records the execution time. This way we can have an objective comparison
+of the performance of the model for various setups.
 
 Benchmark results are tracked in the following Github issue: https://github.com/ggerganov/whisper.cpp/issues/89
 

examples/bench/bench.cpp

@@ -6,9 +6,10 @@
 
 // command-line parameters
 struct whisper_params {
-    int32_t n_threads   = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat
 
-    std::string model     = "models/ggml-base.en.bin";
+    std::string model = "models/ggml-base.en.bin";
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -17,14 +18,14 @@ 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 == "-t" || arg == "--threads") {
-            params.n_threads = std::stoi(argv[++i]);
-        } else if (arg == "-m" || arg == "--model") {
-            params.model = argv[++i];
-        } else if (arg == "-h" || arg == "--help") {
+        if (arg == "-h" || arg == "--help") {
             whisper_print_usage(argc, argv, params);
             exit(0);
-        } else {
+        }
+        else if (arg == "-t" || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
+        else if (arg == "-m" || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w" || arg == "--what")    { params.what     = atoi(argv[++i]); }
+        else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
             exit(0);
@@ -34,27 +35,25 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+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           show this help message and exit\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    model path (default: %s)\n", params.model.c_str());
+    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, "  -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  1 - memcpy\n",                                  "");
+    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
     fprintf(stderr, "\n");
 }
 
-int main(int argc, char ** argv) {
-    whisper_params params;
-
-    if (whisper_params_parse(argc, argv, params) == false) {
-        return 1;
-    }
-
+int whisper_bench_encoder(const whisper_params & params) {
     // whisper init
 
-    struct whisper_context * ctx = whisper_init(params.model.c_str());
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
 
     {
         fprintf(stderr, "\n");
@@ -93,3 +92,22 @@ int main(int argc, char ** argv) {
 
     return 0;
 }
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    int ret = -1;
+
+    switch (params.what) {
+        case 0: ret = whisper_bench_encoder(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;
+    }
+
+    return ret;
+}

examples/command.wasm/CMakeLists.txt

@@ -0,0 +1,50 @@
+#
+# libcommand
+#
+
+set(TARGET libcommand)
+
+add_executable(${TARGET}
+    emscripten.cpp
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE
+    common
+    whisper
+    )
+
+unset(EXTRA_FLAGS)
+
+if (WHISPER_WASM_SINGLE_FILE)
+    set(EXTRA_FLAGS "-s SINGLE_FILE=1")
+    message(STATUS "Embedding WASM inside command.js")
+
+    add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_BINARY_DIR}/bin/libcommand.js
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/command.wasm/command.js
+        )
+endif()
+
+set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
+    --bind \
+    -s USE_PTHREADS=1 \
+    -s PTHREAD_POOL_SIZE=8 \
+    -s INITIAL_MEMORY=1024MB \
+    -s TOTAL_MEMORY=1024MB \
+    -s FORCE_FILESYSTEM=1 \
+    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    ${EXTRA_FLAGS} \
+    ")
+
+#
+# command.wasm
+#
+
+set(TARGET command.wasm)
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/../helpers.js    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/helpers.js @ONLY)

examples/command.wasm/README.md

@@ -0,0 +1,23 @@
+# command.wasm
+
+This is a basic Voice Assistant example that accepts voice commands from the microphone.
+It runs in fully in the browser via WebAseembly.
+
+Online demo: https://whisper.ggerganov.com/command/
+
+Terminal version: [examples/command](/examples/command)
+
+## Build instructions
+
+```bash
+# build using Emscripten (v3.1.2)
+git clone https://github.com/ggerganov/whisper.cpp
+cd whisper.cpp
+mkdir build-em && cd build-em
+emcmake cmake ..
+make -j
+
+# copy the produced page to your HTTP path
+cp bin/command.wasm/*       /path/to/html/
+cp bin/libcommand.worker.js /path/to/html/
+```

examples/command.wasm/emscripten.cpp

@@ -0,0 +1,352 @@
+#include "ggml.h"
+#include "common.h"
+#include "whisper.h"
+
+#include <emscripten.h>
+#include <emscripten/bind.h>
+
+#include <atomic>
+#include <cmath>
+#include <mutex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+constexpr int N_THREAD = 8;
+
+std::vector<struct whisper_context *> g_contexts(4, nullptr);
+
+std::mutex  g_mutex;
+std::thread g_worker;
+
+std::atomic<bool> g_running(false);
+
+std::string g_status        = "";
+std::string g_status_forced = "";
+std::string g_transcribed   = "";
+
+std::vector<float> g_pcmf32;
+
+// 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;
+}
+
+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();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+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 (n_samples > (int) g_pcmf32.size()) {
+        n_take = g_pcmf32.size();
+    } else {
+        n_take = n_samples;
+    }
+
+    audio.resize(n_take);
+    std::copy(g_pcmf32.end() - n_take, g_pcmf32.end(), audio.begin());
+}
+
+void command_main(size_t index) {
+    command_set_status("loading data ...");
+
+    struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
+
+    wparams.n_threads        = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+    wparams.offset_ms        = 0;
+    wparams.translate        = false;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.print_realtime   = false;
+    wparams.print_progress   = false;
+    wparams.print_timestamps = true;
+    wparams.print_special    = false;
+
+    wparams.max_tokens       = 32;
+    wparams.audio_ctx        = 768; // partial encoder context for better performance
+
+    wparams.language         = "en";
+
+    printf("command: using %d threads\n", wparams.n_threads);
+
+    bool have_prompt  = false;
+    bool ask_prompt   = true;
+    bool print_energy = false;
+
+    float prob0 = 0.0f;
+    float prob  = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string k_prompt = "Ok Whisper, start listening for commands.";
+
+    // whisper context
+    auto & ctx = g_contexts[index];
+
+    const int32_t vad_ms     = 2000;
+    const int32_t prompt_ms  = 5000;
+    const int32_t command_ms = 4000;
+
+    const float vad_thold  = 0.1f;
+    const float freq_thold = -1.0f;
+
+    while (g_running) {
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            {
+                char txt[1024];
+                snprintf(txt, sizeof(txt), "Say the following phrase: '%s'", k_prompt.c_str());
+                command_set_status(txt);
+            }
+
+            ask_prompt = false;
+        }
+
+        int64_t t_ms = 0;
+
+        {
+            command_get_audio(vad_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
+
+            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 ...");
+
+                if (!have_prompt) {
+                    command_get_audio(prompt_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
+
+                    const auto txt = ::trim(::command_transcribe(ctx, wparams, pcmf32_cur, prob0, t_ms));
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+
+                    const float sim = similarity(txt, k_prompt);
+
+                    if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                        fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                        ask_prompt = true;
+                    } else {
+                        fprintf(stdout, "\n");
+                        fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                        fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                        fprintf(stdout, "\n");
+
+                        {
+                            char txt[1024];
+                            snprintf(txt, sizeof(txt), "Success! Waiting for voice commands ...");
+                            command_set_status(txt);
+                        }
+
+                        // save the audio for the prompt
+                        pcmf32_prompt = pcmf32_cur;
+                        have_prompt = true;
+                    }
+                } else {
+                    command_get_audio(command_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
+
+                    // prepend the prompt audio
+                    pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+
+                    const auto txt = ::trim(::command_transcribe(ctx, wparams, pcmf32_cur, prob, t_ms));
+
+                    prob = 100.0f*(prob - prob0);
+
+                    fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                    // find the prompt in the text
+                    float best_sim = 0.0f;
+                    size_t best_len = 0;
+                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        const auto prompt = txt.substr(0, n);
+
+                        const float sim = similarity(prompt, k_prompt);
+
+                        //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                        if (sim > best_sim) {
+                            best_sim = sim;
+                            best_len = n;
+                        }
+                    }
+
+                    const std::string command = ::trim(txt.substr(best_len));
+
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    fprintf(stdout, "\n");
+
+                    {
+                        char txt[1024];
+                        snprintf(txt, sizeof(txt), "Command '%s', (t = %d ms)", command.c_str(), (int) t_ms);
+                        command_set_status(txt);
+                    }
+                    {
+                        std::lock_guard<std::mutex> lock(g_mutex);
+                        g_transcribed = command;
+                    }
+                }
+
+                g_pcmf32.clear();
+            }
+        }
+    }
+
+    if (index < g_contexts.size()) {
+        whisper_free(g_contexts[index]);
+        g_contexts[index] = nullptr;
+    }
+}
+
+EMSCRIPTEN_BINDINGS(command) {
+    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
+        for (size_t i = 0; i < g_contexts.size(); ++i) {
+            if (g_contexts[i] == nullptr) {
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                if (g_contexts[i] != nullptr) {
+                    g_running = true;
+                    if (g_worker.joinable()) {
+                        g_worker.join();
+                    }
+                    g_worker = std::thread([i]() {
+                        command_main(i);
+                    });
+
+                    return i + 1;
+                } else {
+                    return (size_t) 0;
+                }
+            }
+        }
+
+        return (size_t) 0;
+    }));
+
+    emscripten::function("free", emscripten::optional_override([](size_t index) {
+        if (g_running) {
+            g_running = false;
+        }
+    }));
+
+    emscripten::function("set_audio", emscripten::optional_override([](size_t index, const emscripten::val & audio) {
+        --index;
+
+        if (index >= g_contexts.size()) {
+            return -1;
+        }
+
+        if (g_contexts[index] == nullptr) {
+            return -2;
+        }
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            const int n = audio["length"].as<int>();
+
+            emscripten::val heap = emscripten::val::module_property("HEAPU8");
+            emscripten::val memory = heap["buffer"];
+
+            g_pcmf32.resize(n);
+
+            emscripten::val memoryView = audio["constructor"].new_(memory, reinterpret_cast<uintptr_t>(g_pcmf32.data()), n);
+            memoryView.call<void>("set", audio);
+        }
+
+        return 0;
+    }));
+
+    emscripten::function("get_transcribed", emscripten::optional_override([]() {
+        std::string transcribed;
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            transcribed = std::move(g_transcribed);
+        }
+
+        return transcribed;
+    }));
+
+    emscripten::function("get_status", emscripten::optional_override([]() {
+        std::string status;
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            status = g_status_forced.empty() ? g_status : g_status_forced;
+        }
+
+        return status;
+    }));
+
+    emscripten::function("set_status", emscripten::optional_override([](const std::string & status) {
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            g_status_forced = status;
+        }
+    }));
+}

examples/command.wasm/index-tmpl.html

@@ -0,0 +1,386 @@
+<!doctype html>
+<html lang="en-us">
+    <head>
+        <title>command : Voice assistant example using Whisper + WebAssembly</title>
+
+        <style>
+            #output {
+                width: 100%;
+                height: 100%;
+                margin: 0 auto;
+                margin-top: 10px;
+                border-left: 0px;
+                border-right: 0px;
+                padding-left: 0px;
+                padding-right: 0px;
+                display: block;
+                background-color: black;
+                color: white;
+                font-size: 10px;
+                font-family: 'Lucida Console', Monaco, monospace;
+                outline: none;
+                white-space: pre;
+                overflow-wrap: normal;
+                overflow-x: scroll;
+            }
+        </style>
+    </head>
+    <body>
+        <div id="main-container">
+            <b>command : Voice assistant example using Whisper + WebAssembly</b>
+
+            <br><br>
+
+            You can find more about this project on <a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/command.wasm">GitHub</a>.
+
+            <br><br>
+
+            <hr>
+
+            Select the model you would like to use, click the "Start" button and follow the instructions.
+
+            <br><br>
+
+            <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>
+
+                <!--
+                    <input type="file" id="file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                -->
+            </div>
+
+            <br>
+
+            <div id="input">
+                <button id="start" onclick="onStart()" disabled>Start</button>
+                <button id="stop"  onclick="onStop()" disabled>Stop</button>
+                <button id="clear" onclick="clearCache()">Clear Cache</button>
+            </div>
+
+            <br>
+
+            <div id="state">
+                Status: <b><span id="state-status">not started</span></b>
+
+                <pre id="state-transcribed">[The recognized voice commands will be displayed here]</pre>
+            </div>
+
+            <hr>
+
+            Debug output:
+            <textarea id="output" rows="20"></textarea>
+
+            <br>
+
+            <b>Troubleshooting</b>
+
+            <br><br>
+
+            The page does some heavy computations, so make sure:
+
+            <ul>
+                <li>To use a modern web browser (e.g. Chrome, Firefox)</li>
+                <li>To use a fast desktop or laptop computer (i.e. not a mobile phone)</li>
+                <li>Your browser supports WASM <a href="https://webassembly.org/roadmap/">Fixed-width SIMD</a></li>
+            </ul>
+
+            <div class="cell-version">
+                <span>
+                    |
+                    Build time: <span class="nav-link">@GIT_DATE@</span> |
+                    Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
+                    Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
+                    <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/command.wasm">Source Code</a> |
+                </span>
+            </div>
+        </div>
+
+        <script type="text/javascript" src="helpers.js"></script>
+        <script type='text/javascript'>
+            // web audio context
+            var context = null;
+
+            // audio data
+            var audio = null;
+            var audio0 = null;
+
+            // the command instance
+            var instance = null;
+
+            // model name
+            var model_whisper = null;
+
+            var Module = {
+                print: printTextarea,
+                printErr: printTextarea,
+                setStatus: function(text) {
+                    printTextarea('js: ' + text);
+                },
+                monitorRunDependencies: function(left) {
+                },
+                preRun: function() {
+                    printTextarea('js: Preparing ...');
+                },
+                postRun: function() {
+                    printTextarea('js: Initialized successfully!');
+                }
+            };
+
+            //
+            // fetch models
+            //
+
+            let dbVersion = 1
+            let dbName    = 'whisper.ggerganov.com';
+            let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
+
+            function storeFS(fname, buf) {
+                // write to WASM file using FS_createDataFile
+                // if the file exists, delete it
+                try {
+                    Module.FS_unlink(fname);
+                } catch (e) {
+                    // ignore
+                }
+
+                Module.FS_createDataFile("/", fname, buf, true, true);
+
+                printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
+
+                document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
+
+                if (model_whisper != null) {
+                    document.getElementById('start').disabled = false;
+                    document.getElementById('stop' ).disabled = true;
+                }
+            }
+
+            function loadWhisper(model) {
+                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',
+                };
+
+                let sizes = {
+                    'tiny.en': 75,
+                    'base.en': 142,
+                };
+
+                let url     = urls[model];
+                let dst     = 'whisper.bin';
+                let size_mb = sizes[model];
+
+                model_whisper = model;
+
+                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
+
+                cbProgress = function(p) {
+                    let el = document.getElementById('fetch-whisper-progress');
+                    el.innerHTML = Math.round(100*p) + '%';
+                };
+
+                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 = '';
+                };
+
+                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
+            }
+
+            //
+            // microphone
+            //
+
+            const kSampleRate = 16000;
+            const kRestartRecording_s = 120;
+            const kIntervalAudio_ms = 250; // pass the recorded audio to the C++ instance at this rate
+
+            var mediaRecorder = null;
+            var doRecording = false;
+            var startTime = 0;
+
+            window.AudioContext = window.AudioContext || window.webkitAudioContext;
+            window.OfflineAudioContext = window.OfflineAudioContext || window.webkitOfflineAudioContext;
+
+            function stopRecording() {
+                Module.set_status("paused");
+                doRecording = false;
+                audio0 = null;
+                audio = null;
+                context = null;
+            }
+
+            function startRecording() {
+                if (!context) {
+                    context = new AudioContext({
+                        sampleRate: kSampleRate,
+                        channelCount: 1,
+                        echoCancellation: false,
+                        autoGainControl:  true,
+                        noiseSuppression: true,
+                    });
+                }
+
+                Module.set_status("");
+
+                document.getElementById('start').disabled = true;
+                document.getElementById('stop').disabled = false;
+
+                doRecording = true;
+                startTime = Date.now();
+
+                var chunks = [];
+                var stream = null;
+
+                navigator.mediaDevices.getUserMedia({audio: true, video: false})
+                    .then(function(s) {
+                        stream = s;
+                        mediaRecorder = new MediaRecorder(stream);
+                        mediaRecorder.ondataavailable = function(e) {
+                            chunks.push(e.data);
+
+                            var blob = new Blob(chunks, { 'type' : 'audio/ogg; codecs=opus' });
+                            var reader = new FileReader();
+
+                            reader.onload = function(event) {
+                                var buf = new Uint8Array(reader.result);
+
+                                if (!context) {
+                                    return;
+                                }
+                                context.decodeAudioData(buf.buffer, function(audioBuffer) {
+                                    var offlineContext = new OfflineAudioContext(audioBuffer.numberOfChannels, audioBuffer.length, audioBuffer.sampleRate);
+                                    var source = offlineContext.createBufferSource();
+                                    source.buffer = audioBuffer;
+                                    source.connect(offlineContext.destination);
+                                    source.start(0);
+
+                                    offlineContext.startRendering().then(function(renderedBuffer) {
+                                        audio = renderedBuffer.getChannelData(0);
+
+                                        //printTextarea('js: audio recorded, size: ' + audio.length + ', old size: ' + (audio0 == null ? 0 : audio0.length));
+
+                                        var audioAll = new Float32Array(audio0 == null ? audio.length : audio0.length + audio.length);
+                                        if (audio0 != null) {
+                                            audioAll.set(audio0, 0);
+                                        }
+                                        audioAll.set(audio, audio0 == null ? 0 : audio0.length);
+
+                                        if (instance) {
+                                            Module.set_audio(instance, audioAll);
+                                        }
+                                    });
+                                }, function(e) {
+                                    audio = null;
+                                });
+                            }
+
+                            reader.readAsArrayBuffer(blob);
+                        };
+
+                        mediaRecorder.onstop = function(e) {
+                            if (doRecording) {
+                                setTimeout(function() {
+                                    startRecording();
+                                });
+                            }
+                        };
+
+                        mediaRecorder.start(kIntervalAudio_ms);
+                    })
+                    .catch(function(err) {
+                        printTextarea('js: error getting audio stream: ' + err);
+                    });
+
+                var interval = setInterval(function() {
+                    if (!doRecording) {
+                        clearInterval(interval);
+                        mediaRecorder.stop();
+                        stream.getTracks().forEach(function(track) {
+                            track.stop();
+                        });
+
+                        document.getElementById('start').disabled = false;
+                        document.getElementById('stop').disabled  = true;
+
+                        mediaRecorder = null;
+                    }
+
+                    // if audio length is more than kRestartRecording_s seconds, restart recording
+                    if (audio != null && audio.length > kSampleRate*kRestartRecording_s) {
+                        if (doRecording) {
+                            //printTextarea('js: restarting recording');
+
+                            clearInterval(interval);
+                            audio0 = audio;
+                            audio = null;
+                            mediaRecorder.stop();
+                            stream.getTracks().forEach(function(track) {
+                                track.stop();
+                            });
+                        }
+                    }
+                }, 100);
+            }
+
+            //
+            // main
+            //
+
+            var nLines = 0;
+            var intervalUpdate = null;
+            var transcribedAll = '';
+
+            function onStart() {
+                if (!instance) {
+                    instance = Module.init('whisper.bin');
+
+                    if (instance) {
+                        printTextarea("js: whisper initialized, instance: " + instance);
+                    }
+                }
+
+                if (!instance) {
+                    printTextarea("js: failed to initialize whisper");
+                    return;
+                }
+
+                startRecording();
+
+                intervalUpdate = setInterval(function() {
+                    var transcribed = Module.get_transcribed();
+
+                    if (transcribed != null && transcribed.length > 1) {
+                        transcribedAll += transcribed + '<br>';
+                        nLines++;
+
+                        // if more than 10 lines, remove the first line
+                        if (nLines > 10) {
+                            var i = transcribedAll.indexOf('<br>');
+                            if (i > 0) {
+                                transcribedAll = transcribedAll.substring(i + 4);
+                                nLines--;
+                            }
+                        }
+                    }
+
+                    document.getElementById('state-status').innerHTML = Module.get_status();
+                    document.getElementById('state-transcribed').innerHTML = transcribedAll;
+                }, 100);
+            }
+
+            function onStop() {
+                stopRecording();
+            }
+
+        </script>
+        <script type="text/javascript" src="command.js"></script>
+    </body>
+</html>

examples/command/CMakeLists.txt

@@ -0,0 +1,9 @@
+if (WHISPER_SUPPORT_SDL2)
+    # command
+    set(TARGET command)
+    add_executable(${TARGET} command.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+endif ()

examples/command/README.md

@@ -0,0 +1,47 @@
+# command
+
+This is a basic Voice Assistant example that accepts voice commands from the microphone.
+More info is available in [issue #171](https://github.com/ggerganov/whisper.cpp/issues/171).
+
+```bash
+# Run with default arguments and small model
+./command -m ./models/ggml-small.en.bin -t 8
+
+# On Raspberry Pi, use tiny or base models + "-ac 768" for better performance
+./command -m ./models/ggml-tiny.en.bin -ac 768 -t 3 -c 0
+```
+
+https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
+
+Web version: [examples/command.wasm](/examples/command.wasm)
+
+## Guided mode
+
+"Guided mode" allows you to specify a list of commands (i.e. strings) and the transcription will be guided to classify your command into one from the list. This can be useful in situations where a device is listening only for a small subset of commands.
+
+Initial tests show that this approach might be extremely efficient in terms of performance, since it integrates very well with the "partial Encoder" idea from #137.
+
+```bash
+# Run in guided mode, the list of allowed commands is in commands.txt
+./command -m ./models/ggml-base.en.bin -cmd ./examples/command/commands.txt
+
+# On Raspberry Pi, in guided mode you can use "-ac 128" for extra performance
+./command -m ./models/ggml-tiny.en.bin -cmd ./examples/command/commands.txt -ac 128 -t 3 -c 0
+```
+
+https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9b8b-aeeb76bee969.mp4
+
+
+## Building
+
+The `command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+make command
+```

examples/command/command.cpp

@@ -0,0 +1,690 @@
+// Voice assistant example
+//
+// Speak short text commands to the microphone.
+// This program will detect your voice command and convert them to text.
+//
+// ref: https://github.com/ggerganov/whisper.cpp/issues/171
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+
+#include <sstream>
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <mutex>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <map>
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t prompt_ms  = 5000;
+    int32_t command_ms = 8000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+
+    std::string language  = "en";
+    std::string model     = "models/ggml-base.en.bin";
+    std::string fname_out;
+    std::string commands;
+    std::string prompt;
+};
+
+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 if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
+        else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else {
+            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, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
+    fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "\n");
+}
+
+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();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    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 = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.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;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    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;
+
+    std::ifstream ifs(fname);
+    if (!ifs.is_open()) {
+        return allowed_commands;
+    }
+
+    std::string line;
+    while (std::getline(ifs, line)) {
+        line = ::trim(line);
+        if (line.empty()) {
+            continue;
+        }
+
+        std::transform(line.begin(), line.end(),line.begin(), ::tolower);
+        allowed_commands.push_back(std::move(line));
+    }
+
+    return allowed_commands;
+}
+
+std::vector<std::string> get_words(const std::string &txt) {
+    std::vector<std::string> words;
+
+    std::istringstream iss(txt);
+    std::string word;
+    while (iss >> word) {
+        words.push_back(word);
+    }
+
+    return words;
+}
+
+// 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) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: guided mode\n", __func__);
+
+    std::vector<std::string> allowed_commands = read_allowed_commands(params.commands);
+
+    if (allowed_commands.empty()) {
+        fprintf(stderr, "%s: error: failed to read allowed commands from '%s'\n", __func__, params.commands.c_str());
+        return 2;
+    }
+
+    int max_len = 0;
+
+    std::vector<std::vector<whisper_token>> allowed_tokens;
+
+    for (const auto & cmd : allowed_commands) {
+        whisper_token tokens[1024];
+        allowed_tokens.emplace_back();
+
+        for (int l = 0; l < (int) cmd.size(); ++l) {
+            // NOTE: very important to add the whitespace !
+            //       the reason is that the first decoded token starts with a whitespace too!
+            std::string ss = std::string(" ") + cmd.substr(0, l + 1);
+
+            const int n = whisper_tokenize(ctx, ss.c_str(), tokens, 1024);
+            if (n < 0) {
+                fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, cmd.c_str());
+                return 3;
+            }
+
+            if (n == 1) {
+                allowed_tokens.back().push_back(tokens[0]);
+            }
+        }
+
+        max_len = std::max(max_len, (int) cmd.size());
+    }
+
+    fprintf(stderr, "%s: allowed commands [ tokens ]:\n", __func__);
+    fprintf(stderr, "\n");
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        fprintf(stderr, "  - \033[1m%-*s\033[0m = [", max_len, allowed_commands[i].c_str());
+        for (const auto & token : allowed_tokens[i]) {
+            fprintf(stderr, " %5d", token);
+        }
+        fprintf(stderr, " ]\n");
+    }
+
+    std::string  k_prompt = "select one from the available words: ";
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        if (i > 0) {
+            k_prompt += ", ";
+        }
+        k_prompt += allowed_commands[i];
+    }
+    k_prompt += ". selected word: ";
+
+    // tokenize prompt
+    std::vector<whisper_token> k_tokens;
+    {
+        k_tokens.resize(1024);
+        const int n = whisper_tokenize(ctx, k_prompt.c_str(), k_tokens.data(), 1024);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to tokenize prompt '%s'\n", __func__, k_prompt.c_str());
+            return 4;
+        }
+        k_tokens.resize(n);
+    }
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: prompt: '%s'\n", __func__, k_prompt.c_str());
+    fprintf(stderr, "%s: tokens: [", __func__);
+    for (const auto & token : k_tokens) {
+        fprintf(stderr, " %d", token);
+    }
+    fprintf(stderr, " ]\n");
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: listening for a command ...\n", __func__);
+    fprintf(stderr, "\n");
+
+    bool is_running  = true;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        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)) {
+            fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+            const auto t_start = std::chrono::high_resolution_clock::now();
+
+            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 = !params.no_timestamps;
+            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;
+
+            wparams.prompt_tokens    = k_tokens.data();
+            wparams.prompt_n_tokens  = k_tokens.size();
+
+            // run the transformer and a single decoding pass
+            if (whisper_full(ctx, wparams, pcmf32_cur.data(), pcmf32_cur.size()) != 0) {
+                fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+                break;
+            }
+
+            // 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;
+
+                double psum = 0.0;
+                for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+                    probs_id.emplace_back(probs[allowed_tokens[i][0]], i);
+                    for (int j = 1; j < (int) allowed_tokens[i].size(); ++j) {
+                        probs_id.back().first += probs[allowed_tokens[i][j]];
+                    }
+                    probs_id.back().first /= allowed_tokens[i].size();
+                    psum += probs_id.back().first;
+                }
+
+                // normalize
+                for (auto & p : probs_id) {
+                    p.first /= psum;
+                }
+
+                // 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;
+                    });
+                }
+
+                // print the commands and the respective probabilities
+                {
+                    fprintf(stdout, "\n");
+                    for (const auto & cmd : probs_id) {
+                        fprintf(stdout, "%s: %s%-*s%s = %f | ", __func__, "\033[1m", max_len, allowed_commands[cmd.second].c_str(), "\033[0m", cmd.first);
+                        for (int token : allowed_tokens[cmd.second]) {
+                            fprintf(stdout, "'%4s' %f ", whisper_token_to_str(ctx, token), probs[token]);
+                        }
+                        fprintf(stdout, "\n");
+                    }
+                }
+
+                // best command
+                {
+                    const auto t_end = std::chrono::high_resolution_clock::now();
+
+                    const float prob = probs_id[0].first;
+                    const int index = probs_id[0].second;
+
+                    fprintf(stdout, "\n");
+                    fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
+                            "\033[1m", allowed_commands[index].c_str(), "\033[0m", prob,
+                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
+                    fprintf(stdout, "\n");
+                }
+            }
+
+            audio.clear();
+        }
+    }
+
+    return 0;
+}
+
+// always-prompt mode
+// transcribe the voice into text after valid prompt
+int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+    bool is_running = true;
+    bool ask_prompt = true;
+
+    float prob = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+
+    const std::string k_prompt = params.prompt;
+
+    const int k_prompt_length = get_words(k_prompt).size();
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: always-prompt mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: The prompt is: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            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;
+
+                // detect the commands
+                audio.get(params.command_ms, pcmf32_cur);
+
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+
+                const auto words = get_words(txt);
+
+                std::string prompt;
+                std::string command;
+
+                for (int i = 0; i < (int) words.size(); ++i) {
+                    if (i < k_prompt_length) {
+                        prompt += words[i] + " ";
+                    } else {
+                        command += words[i] + " ";
+                    }
+                }
+
+                const float sim = similarity(prompt, k_prompt);
+
+                //debug
+                //fprintf(stdout, "command size: %i\n", command_length);
+
+                if ((sim > 0.7f) && (command.size() > 0)) {
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                }
+
+                fprintf(stdout, "\n");
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
+// general-purpose mode
+// freely transcribe the voice into text
+int process_general_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    bool is_running  = true;
+    bool have_prompt = false;
+    bool ask_prompt  = true;
+
+    float prob0 = 0.0f;
+    float prob  = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string k_prompt = "Ok Whisper, start listening for commands.";
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: general-purpose mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            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;
+
+                if (!have_prompt) {
+                    // wait for activation phrase
+                    audio.get(params.prompt_ms, pcmf32_cur);
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+
+                    const float sim = similarity(txt, k_prompt);
+
+                    if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                        fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                        ask_prompt = true;
+                    } else {
+                        fprintf(stdout, "\n");
+                        fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                        fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                        fprintf(stdout, "\n");
+
+                        // save the audio for the prompt
+                        pcmf32_prompt = pcmf32_cur;
+                        have_prompt = true;
+                    }
+                } else {
+                    // we have heard the activation phrase, now detect the commands
+                    audio.get(params.command_ms, pcmf32_cur);
+
+                    // prepend the prompt audio
+                    pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+
+                    prob = 100.0f*(prob - prob0);
+
+                    //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                    // find the prompt in the text
+                    float best_sim = 0.0f;
+                    size_t best_len = 0;
+                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        const auto prompt = txt.substr(0, n);
+
+                        const float sim = similarity(prompt, k_prompt);
+
+                        //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                        if (sim > best_sim) {
+                            best_sim = sim;
+                            best_len = n;
+                        }
+                    }
+
+                    const std::string command = ::trim(txt.substr(best_len));
+
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    fprintf(stdout, "\n");
+                }
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
+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());
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+        if (!whisper_is_multilingual(ctx)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+    // 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();
+
+    // wait for 1 second to avoid any buffered noise
+    std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+    audio.clear();
+
+    int  ret_val = 0;
+
+    if (!params.commands.empty()) {
+        ret_val = process_command_list(ctx, audio, params);
+    } else if (!params.prompt.empty()) {
+        ret_val = always_prompt_transcription(ctx, audio, params);
+    } else {
+        ret_val = process_general_transcription(ctx, audio, params);
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return ret_val;
+}

examples/command/commands.txt

@@ -0,0 +1,9 @@
+enable
+disable
+cat
+dog
+apple
+red
+blue
+green
+lightblue

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 +1,162 @@
+#include "common.h"
+
+// third-party utilities
+// use your favorite implementations
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+
+#include <cmath>
+#include <regex>
+
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+
+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;
+}
+
+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;
+}

examples/common.h

@@ -0,0 +1,40 @@
+#pragma once
+
+// needs to match WHISPER_SAMPLE_RATE
+#define COMMON_SAMPLE_RATE 16000
+
+#include <vector>
+#include <string>
+
+std::string trim(const std::string & s);
+
+std::string replace(
+        const std::string & s,
+        const std::string & from,
+        const std::string & to);
+
+// 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);
+

examples/generate-karaoke.sh

@@ -0,0 +1,60 @@
+#!/bin/bash
+
+# Simple tool to record audio from the microphone and generate a karaoke video
+# Usage:
+#
+#  cd whisper.cpp
+#  make
+#
+#  ./examples/generate-karaoke.sh [model] [step_ms]
+#
+# Press Ctrl+C to stop recording
+#
+
+executable="./main"
+model="base.en"
+model_path="models/ggml-$model.bin"
+
+# require sox and ffmpeg to be installed
+if ! command -v sox &> /dev/null
+then
+    echo "sox could not be found"
+    exit 1
+fi
+
+if ! command -v ffmpeg &> /dev/null
+then
+    echo "ffmpeg could not be found"
+    exit 2
+fi
+
+if [ ! -f "$executable" ]; then
+    echo "'$executable' does not exist. Please build it first."
+    exit 3
+fi
+
+if [ ! -f "$model_path" ]; then
+    echo "'$model_path' does not exist. Please download it first."
+    exit 4
+fi
+
+# record some raw audio
+sox -d rec.wav
+
+# resample to 16kHz
+ffmpeg -y -i ./rec.wav -ar 16000 -ac 1 -c:a pcm_s16le ./rec16.wav > /dev/null 2>&1
+
+# run Whisper
+echo "Processing ..."
+./main -m models/ggml-base.en.bin rec16.wav -owts > /dev/null 2>&1
+
+# generate Karaoke video
+echo "Generating video ..."
+source rec16.wav.wts > /dev/null 2>&1
+
+# play the video
+echo "Playing ./rec16.wav.mp4 ..."
+ffplay -loglevel 0 -autoexit ./rec16.wav.mp4
+
+echo "Done"
+exit 0

examples/helpers.js

@@ -0,0 +1,186 @@
+// Common Javascript functions used by the examples
+
+function convertTypedArray(src, type) {
+    var buffer = new ArrayBuffer(src.byteLength);
+    var baseView = new src.constructor(buffer).set(src);
+    return new type(buffer);
+}
+
+var printTextarea = (function() {
+    var element = document.getElementById('output');
+    if (element) element.value = ''; // clear browser cache
+    return function(text) {
+        if (arguments.length > 1) text = Array.prototype.slice.call(arguments).join(' ');
+        console.log(text);
+        if (element) {
+            element.value += text + "\n";
+            element.scrollTop = element.scrollHeight; // focus on bottom
+        }
+    };
+})();
+
+async function clearCache() {
+    if (confirm('Are you sure you want to clear the cache?\nAll the models will be downloaded again.')) {
+        indexedDB.deleteDatabase(dbName);
+    }
+}
+
+// fetch a remote file from remote URL using the Fetch API
+async function fetchRemote(url, cbProgress, cbPrint) {
+    cbPrint('fetchRemote: downloading with fetch()...');
+
+    const response = await fetch(
+        url,
+        {
+            method: 'GET',
+            headers: {
+                'Content-Type': 'application/octet-stream',
+            },
+        }
+    );
+
+    if (!response.ok) {
+        cbPrint('fetchRemote: failed to fetch ' + url);
+        return;
+    }
+
+    const contentLength = response.headers.get('content-length');
+    const total = parseInt(contentLength, 10);
+    const reader = response.body.getReader();
+
+    var chunks = [];
+    var receivedLength = 0;
+    var progressLast = -1;
+
+    while (true) {
+        const { done, value } = await reader.read();
+
+        if (done) {
+            break;
+        }
+
+        chunks.push(value);
+        receivedLength += value.length;
+
+        if (contentLength) {
+            cbProgress(receivedLength/total);
+
+            var progressCur = Math.round((receivedLength / total) * 10);
+            if (progressCur != progressLast) {
+                cbPrint('fetchRemote: fetching ' + 10*progressCur + '% ...');
+                progressLast = progressCur;
+            }
+        }
+    }
+
+    var position = 0;
+    var chunksAll = new Uint8Array(receivedLength);
+
+    for (var chunk of chunks) {
+        chunksAll.set(chunk, position);
+        position += chunk.length;
+    }
+
+    return chunksAll;
+}
+
+// load remote data
+// - check if the data is already in the IndexedDB
+// - if not, fetch it from the remote URL and store it in the IndexedDB
+function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
+    if (!navigator.storage || !navigator.storage.estimate) {
+        cbPrint('loadRemote: navigator.storage.estimate() is not supported');
+    } else {
+        // query the storage quota and print it
+        navigator.storage.estimate().then(function (estimate) {
+            cbPrint('loadRemote: storage quota: ' + estimate.quota + ' bytes');
+            cbPrint('loadRemote: storage usage: ' + estimate.usage + ' bytes');
+        });
+    }
+
+    // check if the data is already in the IndexedDB
+    var rq = indexedDB.open(dbName, dbVersion);
+
+    rq.onupgradeneeded = function (event) {
+        var db = event.target.result;
+        if (db.version == 1) {
+            var os = db.createObjectStore('models', { autoIncrement: false });
+            cbPrint('loadRemote: created IndexedDB ' + db.name + ' version ' + db.version);
+        } else {
+            // clear the database
+            var os = event.currentTarget.transaction.objectStore('models');
+            os.clear();
+            cbPrint('loadRemote: cleared IndexedDB ' + db.name + ' version ' + db.version);
+        }
+    };
+
+    rq.onsuccess = function (event) {
+        var db = event.target.result;
+        var tx = db.transaction(['models'], 'readonly');
+        var os = tx.objectStore('models');
+        var rq = os.get(url);
+
+        rq.onsuccess = function (event) {
+            if (rq.result) {
+                cbPrint('loadRemote: "' + url + '" is already in the IndexedDB');
+                cbReady(dst, rq.result);
+            } else {
+                // data is not in the IndexedDB
+                cbPrint('loadRemote: "' + url + '" is not in the IndexedDB');
+
+                // alert and ask the user to confirm
+                if (!confirm(
+                    'You are about to download ' + size_mb + ' MB of data.\n' +
+                    'The model data will be cached in the browser for future use.\n\n' +
+                    'Press OK to continue.')) {
+                    cbCancel();
+                    return;
+                }
+
+                fetchRemote(url, cbProgress, cbPrint).then(function (data) {
+                    if (data) {
+                        // store the data in the IndexedDB
+                        var rq = indexedDB.open(dbName, dbVersion);
+                        rq.onsuccess = function (event) {
+                            var db = event.target.result;
+                            var tx = db.transaction(['models'], 'readwrite');
+                            var os = tx.objectStore('models');
+                            var rq = os.put(data, url);
+
+                            rq.onsuccess = function (event) {
+                                cbPrint('loadRemote: "' + url + '" stored in the IndexedDB');
+                                cbReady(dst, data);
+                            };
+
+                            rq.onerror = function (event) {
+                                cbPrint('loadRemote: failed to store "' + url + '" in the IndexedDB');
+                                cbCancel();
+                            };
+                        };
+                    }
+                });
+            }
+        };
+
+        rq.onerror = function (event) {
+            cbPrint('loadRemote: failed to get data from the IndexedDB');
+            cbCancel();
+        };
+    };
+
+    rq.onerror = function (event) {
+        cbPrint('loadRemote: failed to open IndexedDB');
+        cbCancel();
+    };
+
+    rq.onblocked = function (event) {
+        cbPrint('loadRemote: failed to open IndexedDB: blocked');
+        cbCancel();
+    };
+
+    rq.onabort = function (event) {
+        cbPrint('loadRemote: failed to open IndexedDB: abort');
+
+    };
+}
+

examples/livestream.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+#
+# Transcribe audio livestream by feeding ffmpeg output to whisper.cpp at regular intervals
+# Idea by @semiformal-net
+# ref: https://github.com/ggerganov/whisper.cpp/issues/185
+#
+
+set -eo pipefail
+
+url="http://a.files.bbci.co.uk/media/live/manifesto/audio/simulcast/hls/nonuk/sbr_low/ak/bbc_world_service.m3u8"
+fmt=aac # the audio format extension of the stream (TODO: auto detect)
+step_s=30
+model="base.en"
+
+check_requirements()
+{
+    if ! command -v ./main &>/dev/null; then
+        echo "whisper.cpp main executable is required (make)"
+        exit 1
+    fi
+
+    if ! command -v ffmpeg &>/dev/null; then
+        echo "ffmpeg is required (https://ffmpeg.org)"
+        exit 1
+    fi
+}
+
+check_requirements
+
+
+if [ -z "$1" ]; then
+    echo "Usage: $0 stream_url [step_s] [model]"
+    echo ""
+    echo "  Example:"
+    echo "    $0 $url $step_s $model"
+    echo ""
+    echo "No url specified, using default: $url"
+else
+    url="$1"
+fi
+
+if [ -n "$2" ]; then
+    step_s="$2"
+fi
+
+if [ -n "$3" ]; then
+    model="$3"
+fi
+
+# Whisper models
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+running=1
+
+trap "running=0" SIGINT SIGTERM
+
+printf "[+] Transcribing stream with model '$model', step_s $step_s (press Ctrl+C to stop):\n\n"
+
+# continuous stream in native fmt (this file will grow forever!)
+ffmpeg -loglevel quiet -y -re -probesize 32 -i $url -c copy /tmp/whisper-live0.${fmt} &
+if [ $? -ne 0 ]; then
+    printf "Error: ffmpeg failed to capture audio stream\n"
+    exit 1
+fi
+
+printf "Buffering audio. Please wait...\n\n"
+sleep $(($step_s))
+
+# do not stop script on error
+set +e
+
+i=0
+SECONDS=0
+while [ $running -eq 1 ]; do
+    # extract the next piece from the main file above and transcode to wav. -ss sets start time and nudges it by -0.5s to catch missing words (??)
+    err=1
+    while [ $err -ne 0 ]; do
+        if [ $i -gt 0 ]; then
+            ffmpeg -loglevel quiet -v error -noaccurate_seek -i /tmp/whisper-live0.${fmt} -y -ar 16000 -ac 1 -c:a pcm_s16le -ss $(($i*$step_s-1)).5 -t $step_s /tmp/whisper-live.wav 2> /tmp/whisper-live.err
+        else
+            ffmpeg -loglevel quiet -v error -noaccurate_seek -i /tmp/whisper-live0.${fmt} -y -ar 16000 -ac 1 -c:a pcm_s16le -ss $(($i*$step_s)) -t $step_s /tmp/whisper-live.wav 2> /tmp/whisper-live.err
+        fi
+        err=$(cat /tmp/whisper-live.err | wc -l)
+    done
+
+    ./main -t 8 -m ./models/ggml-${model}.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
+
+    while [ $SECONDS -lt $((($i+1)*$step_s)) ]; do
+        sleep 1
+    done
+    ((i=i+1))
+done
+
+killall -v ffmpeg
+killall -v main

examples/main/CMakeLists.txt

@@ -1,3 +1,6 @@
 set(TARGET main)
 add_executable(${TARGET} main.cpp)
-target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/main/README.md

@@ -6,29 +6,39 @@ It can be used as a reference for using the `whisper.cpp` library in other proje
 ```
 ./main -h
 
-usage: ./bin/main [options] file0.wav file1.wav ...
-
-  -h,       --help           show this help message and exit
-  -s SEED,  --seed SEED      RNG seed (default: -1)
-  -t N,     --threads N      number of threads to use during computation (default: 4)
-  -p N,     --processors N   number of processors to use during computation (default: 1)
-  -ot N,    --offset-t N     time offset in milliseconds (default: 0)
-  -on N,    --offset-n N     segment index offset (default: 0)
-  -mc N,    --max-context N  maximum number of text context tokens to store (default: max)
-  -ml N,    --max-len N      maximum segment length in characters (default: 0)
-  -wt N,    --word-thold N   word timestamp probability threshold (default: 0.010000)
-  -v,       --verbose        verbose output
-            --translate      translate from source language to english
-  -otxt,    --output-txt     output result in a text file
-  -ovtt,    --output-vtt     output result in a vtt file
-  -osrt,    --output-srt     output result in a srt file
-  -owts,    --output-words   output script for generating karaoke video
-  -ps,      --print_special  print special tokens
-  -pc,      --print_colors   print colors
-  -nt,      --no_timestamps  do not print timestamps
-  -l LANG,  --language LANG  spoken language (default: en)
-  -m FNAME, --model FNAME    model path (default: models/ggml-base.en.bin)
-  -f FNAME, --file FNAME     input WAV file path
-  -h,       --help           show this help message and exit
+usage: ./main [options] file0.wav file1.wav ...
 
+options:
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
+  -tr,       --translate         [false  ] translate from source language to english
+  -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
+  -ovtt,     --output-vtt        [false  ] output result in a vtt file
+  -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
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [true   ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -f FNAME,  --file FNAME        [       ] input WAV file path
 ```

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 +8,7 @@
 #include <string>
 #include <thread>
 #include <vector>
+#include <cstring>
 
 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
@@ -36,6 +34,10 @@ std::string to_timestamp(int64_t t, bool comma = false) {
     return std::string(buf);
 }
 
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
 // helper function to replace substrings
 void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     for (size_t pos = 0; ; pos += replace.length()) {
@@ -48,30 +50,44 @@ void replace_all(std::string & s, const std::string & search, const std::string
 
 // command-line parameters
 struct whisper_params {
-    int32_t seed         = -1; // RNG seed, not used currently
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors = 1;
-    int32_t offset_t_ms  = 0;
-    int32_t offset_n     = 0;
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
     int32_t max_context  = -1;
-    int32_t max_len      = 0;
+    int32_t max_len      =  0;
+    int32_t best_of      =  2;
+    int32_t beam_size    = -1;
 
-    float word_thold = 0.01f;
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;
 
-    bool verbose              = false;
-    bool translate            = false;
-    bool output_txt           = false;
-    bool output_vtt           = false;
-    bool output_srt           = false;
-    bool output_wts           = false;
-    bool print_special_tokens = false;
-    bool print_colors         = false;
-    bool no_timestamps        = false;
+    bool speed_up       = false;
+    bool translate      = false;
+    bool diarize        = false;
+    bool split_on_word  = false;
+    bool no_fallback    = false;
+    bool output_txt     = false;
+    bool output_vtt     = false;
+    bool output_srt     = false;
+    bool output_wts     = false;
+    bool output_csv     = false;
+    bool output_jsn     = false;
+    bool output_lrc     = false;
+    bool print_special  = false;
+    bool print_colors   = false;
+    bool print_progress = false;
+    bool no_timestamps  = false;
 
-    std::string language  = "en";
-    std::string model     = "models/ggml-base.en.bin";
+    std::string language = "en";
+    std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
+    std::vector<std::string> fname_out = {};
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -80,60 +96,55 @@ 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;
         }
 
-        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 == "-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 == "-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 == "-wt" || arg == "--word-thold") {
-            params.word_thold = std::stof(argv[++i]);
-        } else if (arg == "-v" || arg == "--verbose") {
-            params.verbose = true;
-        } else if (arg == "--translate") {
-            params.translate = true;
-        } else if (arg == "-l" || arg == "--language") {
-            params.language = argv[++i];
-            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);
-            }
-        } 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 == "-ps" || arg == "--print_special") {
-            params.print_special_tokens = true;
-        } else if (arg == "-pc" || arg == "--print_colors") {
-            params.print_colors = true;
-        } else if (arg == "-nt" || arg == "--no_timestamps") {
-            params.no_timestamps = true;
-        } else if (arg == "-m" || arg == "--model") {
-            params.model = argv[++i];
-        } else if (arg == "-f" || arg == "--file") {
-            params.fname_inp.push_back(argv[++i]);
-        } else if (arg == "-h" || arg == "--help") {
+        if (arg == "-h" || arg == "--help") {
             whisper_print_usage(argc, argv, params);
             exit(0);
-        } else {
+        }
+        else if (arg == "-t"    || arg == "--threads")        { params.n_threads      = std::stoi(argv[++i]); }
+        else if (arg == "-p"    || arg == "--processors")     { params.n_processors   = std::stoi(argv[++i]); }
+        else if (arg == "-ot"   || arg == "--offset-t")       { params.offset_t_ms    = std::stoi(argv[++i]); }
+        else if (arg == "-on"   || arg == "--offset-n")       { params.offset_n       = std::stoi(argv[++i]); }
+        else if (arg == "-d"    || arg == "--duration")       { params.duration_ms    = std::stoi(argv[++i]); }
+        else if (arg == "-mc"   || arg == "--max-context")    { params.max_context    = std::stoi(argv[++i]); }
+        else if (arg == "-ml"   || arg == "--max-len")        { params.max_len        = std::stoi(argv[++i]); }
+        else if (arg == "-bo"   || arg == "--best-of")        { params.best_of        = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")      { params.beam_size      = std::stoi(argv[++i]); }
+        else if (arg == "-wt"   || arg == "--word-thold")     { params.word_thold     = std::stof(argv[++i]); }
+        else if (arg == "-et"   || arg == "--entropy-thold")  { params.entropy_thold  = std::stof(argv[++i]); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")  { params.logprob_thold  = std::stof(argv[++i]); }
+        else if (arg == "-su"   || arg == "--speed-up")       { params.speed_up       = true; }
+        else if (arg == "-tr"   || arg == "--translate")      { params.translate      = true; }
+        else if (arg == "-di"   || arg == "--diarize")        { params.diarize        = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")  { params.split_on_word  = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")    { params.no_fallback    = true; }
+        else if (arg == "-otxt" || arg == "--output-txt")     { params.output_txt     = true; }
+        else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
+        else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
+        else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
+        else if (arg == "-olrc" || arg == "--output-lrc")     { params.output_lrc     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
+        else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
+        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
+        else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
+        else if (arg == "-pc"   || arg == "--print-colors")   { params.print_colors   = true; }
+        else if (arg == "-pp"   || arg == "--print-progress") { params.print_progress = true; }
+        else if (arg == "-nt"   || arg == "--no-timestamps")  { params.no_timestamps  = true; }
+        else if (arg == "-l"    || arg == "--language")       { params.language       = argv[++i]; }
+        else if (                  arg == "--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 {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
             exit(0);
@@ -143,97 +154,136 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
     fprintf(stderr, "\n");
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\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, "  -p N,     --processors N   number of processors to use during computation (default: %d)\n", params.n_processors);
-    fprintf(stderr, "  -ot N,    --offset-t N     time offset in milliseconds (default: %d)\n", params.offset_t_ms);
-    fprintf(stderr, "  -on N,    --offset-n N     segment index offset (default: %d)\n", params.offset_n);
-    fprintf(stderr, "  -mc N,    --max-context N  maximum number of text context tokens to store (default: max)\n");
-    fprintf(stderr, "  -ml N,    --max-len N      maximum segment length in characters (default: %d)\n", params.max_len);
-    fprintf(stderr, "  -wt N,    --word-thold N   word timestamp probability threshold (default: %f)\n", params.word_thold);
-    fprintf(stderr, "  -v,       --verbose        verbose output\n");
-    fprintf(stderr, "            --translate      translate from source language to english\n");
-    fprintf(stderr, "  -otxt,    --output-txt     output result in a text file\n");
-    fprintf(stderr, "  -ovtt,    --output-vtt     output result in a vtt file\n");
-    fprintf(stderr, "  -osrt,    --output-srt     output result in a srt file\n");
-    fprintf(stderr, "  -owts,    --output-words   output script for generating karaoke video\n");
-    fprintf(stderr, "  -ps,      --print_special  print special tokens\n");
-    fprintf(stderr, "  -pc,      --print_colors   print colors\n");
-    fprintf(stderr, "  -nt,      --no_timestamps  do not print timestamps\n");
-    fprintf(stderr, "  -l LANG,  --language LANG  spoken language (default: %s)\n", params.language.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME    model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME     input WAV file path\n");
+    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -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, "  -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, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    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, "\n");
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
-    const whisper_params & params = *(whisper_params *) user_data;
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+};
+
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
     const int n_segments = whisper_full_n_segments(ctx);
 
+    std::string speaker = "";
+
+    int64_t t0 = 0;
+    int64_t t1 = 0;
+
     // print the last n_new segments
     const int s0 = n_segments - n_new;
+
     if (s0 == 0) {
         printf("\n");
     }
 
     for (int i = s0; i < n_segments; i++) {
-        if (params.no_timestamps) {
-            if (params.print_colors) {
-                for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-                    if (params.print_special_tokens == false) {
-                        const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-                        if (id >= whisper_token_eot(ctx)) {
-                            continue;
-                        }
-                    }
-
-                    const char * text = whisper_full_get_token_text(ctx, i, j);
-                    const float  p    = whisper_full_get_token_p   (ctx, i, j);
-
-                    const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));
-
-                    printf("%s%s%s", k_colors[col].c_str(), text, "\033[0m");
-                }
-            } else {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-                printf("%s", text);
-            }
-            fflush(stdout);
-        } else {
-            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-
-            if (params.print_colors) {
-                printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
-                for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-                    if (params.print_special_tokens == false) {
-                        const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-                        if (id >= whisper_token_eot(ctx)) {
-                            continue;
-                        }
-                    }
-
-                    const char * text = whisper_full_get_token_text(ctx, i, j);
-                    const float  p    = whisper_full_get_token_p   (ctx, i, j);
-
-                    const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));
-
-                    printf("%s%s%s", k_colors[col].c_str(), text, "\033[0m");
-                }
-                printf("\n");
-            } else {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-
-                printf("[%s --> %s]  %s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), text);
-            }
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
         }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            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());
+        }
+
+        if (params.print_colors) {
+            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
+                if (params.print_special == false) {
+                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
+                    if (id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+                }
+
+                const char * text = whisper_full_get_token_text(ctx, i, j);
+                const float  p    = whisper_full_get_token_p   (ctx, i, j);
+
+                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
+
+                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
+            }
+        } else {
+            const char * text = whisper_full_get_segment_text(ctx, i);
+
+            printf("%s%s", speaker.c_str(), text);
+        }
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+
+        fflush(stdout);
     }
 }
 
@@ -249,7 +299,7 @@ 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;
+        fout << text << "\n";
     }
 
     return true;
@@ -259,7 +309,7 @@ bool output_vtt(struct whisper_context * ctx, const char * fname) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
         fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
-        return 9;
+        return false;
     }
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@@ -302,6 +352,187 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
+bool output_csv(struct whisper_context * ctx, const char * fname) {
+    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);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,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);
+
+        //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";
+    }
+
+    return true;
+}
+
+char *escape_double_quotes(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] == '"') {
+            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] == '"') {
+            escaped[pos++] = '\\';
+            escaped[pos++] = '"';
+        } else {
+            escaped[pos++] = str[i];
+        }
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+    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 = false) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "},\n");
+    };
+
+    auto start_obj = [&](const char *name = nullptr) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end = false) {
+        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 = false) {
+        start_value(name);
+        char * val_escaped = escape_double_quotes(val);
+        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
+        free(val_escaped);
+    };
+
+    auto end_value = [&](bool end = false) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    auto value_i = [&](const char *name, const int64_t val, bool end = false) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end = false) {
+        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();
+        value_s("systeminfo", whisper_print_system_info());
+        start_obj("model");
+            value_s("type", whisper_model_type_readable(ctx));
+            value_b("multilingual", whisper_is_multilingual(ctx));
+            value_i("vocab", whisper_model_n_vocab(ctx));
+            start_obj("audio");
+                value_i("ctx", whisper_model_n_audio_ctx(ctx));
+                value_i("state", whisper_model_n_audio_state(ctx));
+                value_i("head", whisper_model_n_audio_head(ctx));
+                value_i("layer", whisper_model_n_audio_layer(ctx), true);
+            end_obj();
+            start_obj("text");
+                value_i("ctx", whisper_model_n_text_ctx(ctx));
+                value_i("state", whisper_model_n_text_state(ctx));
+                value_i("head", whisper_model_n_text_head(ctx));
+                value_i("layer", whisper_model_n_text_layer(ctx), true);
+            end_obj();
+            value_i("mels", whisper_model_n_mels(ctx));
+            value_i("f16", whisper_model_f16(ctx), true);
+        end_obj();
+        start_obj("params");
+            value_s("model", params.model.c_str());
+            value_s("language", params.language.c_str());
+            value_b("translate", params.translate, true);
+        end_obj();
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        end_obj();
+        start_arr("transcription");
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+                start_obj();
+                    start_obj("timestamps");
+                        value_s("from", to_timestamp(t0, true).c_str());
+                        value_s("to", to_timestamp(t1, true).c_str(), true);
+                    end_obj();
+                    start_obj("offsets");
+                        value_i("from", t0 * 10);
+                        value_i("to", t1 * 10, true);
+                    end_obj();
+                    value_s("text", text, true);
+                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
@@ -310,8 +541,13 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
 
     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";
@@ -354,7 +590,6 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
             txt_ul = "\\ \\ ";
 
             {
-                int ncnt = 0;
                 for (int k = 0; k < n; ++k) {
                     const auto & token2 = tokens[k];
 
@@ -378,13 +613,11 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
                             txt_ul += "\\ ";
                         }
                     }
-
-                    ncnt += txt.size();
                 }
 
-                ::replace_all(txt_bg, "'", "’");
+                ::replace_all(txt_bg, "'", "\u2019");
                 ::replace_all(txt_bg, "\"", "\\\"");
-                ::replace_all(txt_fg, "'", "’");
+                ::replace_all(txt_fg, "'", "\u2019");
                 ::replace_all(txt_fg, "\"", "\\\"");
             }
 
@@ -417,6 +650,39 @@ 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) {
+
+    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);
+
+        fout <<  '[' << timestamp_lrc << ']' << text << "\n";
+    }
+
+    return true;
+}
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -424,19 +690,21 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.seed < 0) {
-        params.seed = time(NULL);
-    }
-
     if (params.fname_inp.empty()) {
         fprintf(stderr, "error: no input files specified\n");
         whisper_print_usage(argc, argv, params);
         return 2;
     }
 
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
     // whisper init
 
-    struct whisper_context * ctx = whisper_init(params.model.c_str());
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
 
     if (ctx == nullptr) {
         fprintf(stderr, "error: failed to initialize whisper context\n");
@@ -445,50 +713,14 @@ int main(int argc, char ** argv) {
 
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
+		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
 
-        // WAV input
-        std::vector<float> pcmf32;
-        {
-            drwav wav;
-            if (!drwav_init_file(&wav, fname_inp.c_str(), NULL)) {
-                fprintf(stderr, "%s: failed to open WAV file '%s' - check your input\n", argv[0], fname_inp.c_str());
-                whisper_print_usage(argc, argv, {});
-                return 4;
-            }
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-            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 5;
-            }
-
-            if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16 kHz\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (wav.bitsPerSample != 16) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", argv[0], fname_inp.c_str());
-                return 7;
-            }
-
-            int n = wav.totalPCMFrameCount;
-
-            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 (int i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[i])/32768.0f;
-                }
-            } else {
-                for (int i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-                }
-            }
+        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
@@ -518,34 +750,63 @@ int main(int argc, char ** argv) {
             fprintf(stderr, "\n");
         }
 
-
         // run the inference
         {
             whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
-            wparams.print_realtime       = false;
-            wparams.print_progress       = false;
-            wparams.print_timestamps     = !params.no_timestamps;
-            wparams.print_special_tokens = params.print_special_tokens;
-            wparams.translate            = params.translate;
-            wparams.language             = params.language.c_str();
-            wparams.n_threads            = params.n_threads;
-            wparams.n_max_text_ctx       = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
-            wparams.offset_ms            = params.offset_t_ms;
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
 
-            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.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.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.initial_prompt   = params.prompt.c_str();
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+
+            whisper_print_user_data user_data = { &params, &pcmf32s };
 
             // this callback is called on each new segment
             if (!wparams.print_realtime) {
                 wparams.new_segment_callback           = whisper_print_segment_callback;
-                wparams.new_segment_callback_user_data = &params;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            // example for abort mechanism
+            // in this example, we do not abort the processing, but we could if the flag is set to true
+            // the callback is called before every encoder run - if it returns false, the processing is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
             }
 
             if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
-                return 8;
+                return 10;
             }
         }
 
@@ -555,27 +816,45 @@ int main(int argc, char ** argv) {
 
             // output to text file
             if (params.output_txt) {
-                const auto fname_txt = fname_inp + ".txt";
+                const auto fname_txt = fname_out + ".txt";
                 output_txt(ctx, fname_txt.c_str());
             }
 
             // output to VTT file
             if (params.output_vtt) {
-                const auto fname_vtt = fname_inp + ".vtt";
+                const auto fname_vtt = fname_out + ".vtt";
                 output_vtt(ctx, fname_vtt.c_str());
             }
 
             // output to SRT file
             if (params.output_srt) {
-                const auto fname_srt = fname_inp + ".srt";
+                const auto fname_srt = fname_out + ".srt";
                 output_srt(ctx, fname_srt.c_str(), params);
             }
 
             // output to WTS file
             if (params.output_wts) {
-                const auto fname_wts = fname_inp + ".wts";
+                const auto fname_wts = fname_out + ".wts";
                 output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);
             }
+
+            // output to CSV file
+            if (params.output_csv) {
+                const auto fname_csv = fname_out + ".csv";
+                output_csv(ctx, fname_csv.c_str());
+            }
+
+            // output to JSON file
+            if (params.output_jsn) {
+                const auto fname_jsn = fname_out + ".json";
+                output_json(ctx, fname_jsn.c_str(), params);
+            }
+
+            // output to LRC file
+            if (params.output_lrc) {
+                const auto fname_lrc = fname_out + ".lrc";
+                output_lrc(ctx, fname_lrc.c_str());
+            }
         }
     }
 

examples/stream.wasm/CMakeLists.txt

@@ -0,0 +1,49 @@
+#
+# libstream
+#
+
+set(TARGET libstream)
+
+add_executable(${TARGET}
+    emscripten.cpp
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE
+    whisper
+    )
+
+unset(EXTRA_FLAGS)
+
+if (WHISPER_WASM_SINGLE_FILE)
+    set(EXTRA_FLAGS "-s SINGLE_FILE=1")
+    message(STATUS "Embedding WASM inside stream.js")
+
+    add_custom_command(
+        TARGET ${TARGET} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy
+        ${CMAKE_BINARY_DIR}/bin/libstream.js
+        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/stream.wasm/stream.js
+        )
+endif()
+
+set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
+    --bind \
+    -s USE_PTHREADS=1 \
+    -s PTHREAD_POOL_SIZE=8 \
+    -s INITIAL_MEMORY=1024MB \
+    -s TOTAL_MEMORY=1024MB \
+    -s FORCE_FILESYSTEM=1 \
+    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
+    ${EXTRA_FLAGS} \
+    ")
+
+#
+# stream.wasm
+#
+
+set(TARGET stream.wasm)
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/../helpers.js    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/helpers.js @ONLY)

examples/stream.wasm/README.md

@@ -0,0 +1,20 @@
+# stream.wasm
+
+Real-time transcription in the browser using WebAssembly
+
+Online demo: https://whisper.ggerganov.com/stream/
+
+## Build instructions
+
+```bash
+# build using Emscripten (v3.1.2)
+git clone https://github.com/ggerganov/whisper.cpp
+cd whisper.cpp
+mkdir build-em && cd build-em
+emcmake cmake ..
+make -j
+
+# copy the produced page to your HTTP path
+cp bin/stream.wasm/*       /path/to/html/
+cp bin/libstream.worker.js /path/to/html/
+```

examples/stream.wasm/emscripten.cpp

@@ -0,0 +1,216 @@
+#include "ggml.h"
+#include "whisper.h"
+
+#include <emscripten.h>
+#include <emscripten/bind.h>
+
+#include <atomic>
+#include <cmath>
+#include <mutex>
+#include <string>
+#include <thread>
+#include <vector>
+
+constexpr int N_THREAD = 8;
+
+std::vector<struct whisper_context *> g_contexts(4, nullptr);
+
+std::mutex g_mutex;
+std::thread g_worker;
+
+std::atomic<bool> g_running(false);
+
+std::string g_status        = "";
+std::string g_status_forced = "";
+std::string g_transcribed   = "";
+
+std::vector<float> g_pcmf32;
+
+void stream_set_status(const std::string & status) {
+    std::lock_guard<std::mutex> lock(g_mutex);
+    g_status = status;
+}
+
+void stream_main(size_t index) {
+    stream_set_status("loading data ...");
+
+    struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
+
+    wparams.n_threads        = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
+    wparams.offset_ms        = 0;
+    wparams.translate        = false;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.print_realtime   = false;
+    wparams.print_progress   = false;
+    wparams.print_timestamps = true;
+    wparams.print_special    = false;
+
+    wparams.max_tokens       = 32;
+    wparams.audio_ctx        = 768; // partial encoder context for better performance
+
+    // disable temperature fallback
+    wparams.temperature_inc  = -1.0f;
+
+    wparams.language         = "en";
+
+    printf("stream: using %d threads\n", wparams.n_threads);
+
+    std::vector<float> pcmf32;
+
+    // whisper context
+    auto & ctx = g_contexts[index];
+
+    // 5 seconds interval
+    const int64_t window_samples = 5*WHISPER_SAMPLE_RATE;
+
+    while (g_running) {
+        stream_set_status("waiting for audio ...");
+
+        {
+            std::unique_lock<std::mutex> lock(g_mutex);
+
+            if (g_pcmf32.size() < 1024) {
+                lock.unlock();
+
+                std::this_thread::sleep_for(std::chrono::milliseconds(10));
+
+                continue;
+            }
+
+            pcmf32 = std::vector<float>(g_pcmf32.end() - std::min((int64_t) g_pcmf32.size(), window_samples), g_pcmf32.end());
+            g_pcmf32.clear();
+        }
+
+        {
+            const auto t_start = std::chrono::high_resolution_clock::now();
+
+            stream_set_status("running whisper ...");
+
+            int ret = whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size());
+            if (ret != 0) {
+                printf("whisper_full() failed: %d\n", ret);
+                break;
+            }
+
+            const auto t_end = std::chrono::high_resolution_clock::now();
+
+            printf("stream: whisper_full() returned %d in %f seconds\n", ret, std::chrono::duration<double>(t_end - t_start).count());
+        }
+
+        {
+            std::string text_heard;
+
+            {
+                const int n_segments = whisper_full_n_segments(ctx);
+                for (int i = n_segments - 1; 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);
+
+                    printf("transcribed: %s\n", text);
+
+                    text_heard += text;
+                }
+            }
+
+            {
+                std::lock_guard<std::mutex> lock(g_mutex);
+                g_transcribed = text_heard;
+            }
+        }
+    }
+
+    if (index < g_contexts.size()) {
+        whisper_free(g_contexts[index]);
+        g_contexts[index] = nullptr;
+    }
+}
+
+EMSCRIPTEN_BINDINGS(stream) {
+    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
+        for (size_t i = 0; i < g_contexts.size(); ++i) {
+            if (g_contexts[i] == nullptr) {
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
+                if (g_contexts[i] != nullptr) {
+                    g_running = true;
+                    if (g_worker.joinable()) {
+                        g_worker.join();
+                    }
+                    g_worker = std::thread([i]() {
+                        stream_main(i);
+                    });
+
+                    return i + 1;
+                } else {
+                    return (size_t) 0;
+                }
+            }
+        }
+
+        return (size_t) 0;
+    }));
+
+    emscripten::function("free", emscripten::optional_override([](size_t index) {
+        if (g_running) {
+            g_running = false;
+        }
+    }));
+
+    emscripten::function("set_audio", emscripten::optional_override([](size_t index, const emscripten::val & audio) {
+        --index;
+
+        if (index >= g_contexts.size()) {
+            return -1;
+        }
+
+        if (g_contexts[index] == nullptr) {
+            return -2;
+        }
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            const int n = audio["length"].as<int>();
+
+            emscripten::val heap = emscripten::val::module_property("HEAPU8");
+            emscripten::val memory = heap["buffer"];
+
+            g_pcmf32.resize(n);
+
+            emscripten::val memoryView = audio["constructor"].new_(memory, reinterpret_cast<uintptr_t>(g_pcmf32.data()), n);
+            memoryView.call<void>("set", audio);
+        }
+
+        return 0;
+    }));
+
+    emscripten::function("get_transcribed", emscripten::optional_override([]() {
+        std::string transcribed;
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            transcribed = std::move(g_transcribed);
+        }
+
+        return transcribed;
+    }));
+
+    emscripten::function("get_status", emscripten::optional_override([]() {
+        std::string status;
+
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            status = g_status_forced.empty() ? g_status : g_status_forced;
+        }
+
+        return status;
+    }));
+
+    emscripten::function("set_status", emscripten::optional_override([](const std::string & status) {
+        {
+            std::lock_guard<std::mutex> lock(g_mutex);
+            g_status_forced = status;
+        }
+    }));
+}

examples/stream.wasm/index-tmpl.html

@@ -0,0 +1,386 @@
+<!doctype html>
+<html lang="en-us">
+    <head>
+        <title>stream : Real-time Whisper transcription in WebAssembly</title>
+
+        <style>
+            #output {
+                width: 100%;
+                height: 100%;
+                margin: 0 auto;
+                margin-top: 10px;
+                border-left: 0px;
+                border-right: 0px;
+                padding-left: 0px;
+                padding-right: 0px;
+                display: block;
+                background-color: black;
+                color: white;
+                font-size: 10px;
+                font-family: 'Lucida Console', Monaco, monospace;
+                outline: none;
+                white-space: pre;
+                overflow-wrap: normal;
+                overflow-x: scroll;
+            }
+        </style>
+    </head>
+    <body>
+        <div id="main-container">
+            <b>stream : Real-time Whisper transcription in WebAssembly</b>
+
+            <br><br>
+
+            You can find more about this project on <a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/stream.wasm">GitHub</a>.
+
+            <br><br>
+
+            <hr>
+
+            Select the model you would like to use, click the "Start" button and start speaking
+
+            <br><br>
+
+            <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>
+
+                <!--
+                    <input type="file" id="file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                -->
+            </div>
+
+            <br>
+
+            <div id="input">
+                <button id="start"  onclick="onStart()" disabled>Start</button>
+                <button id="stop"   onclick="onStop()" disabled>Stop</button>
+                <button id="clear"  onclick="clearCache()">Clear Cache</button>
+            </div>
+
+            <br>
+
+            <div id="state">
+                Status: <b><span id="state-status">not started</span></b>
+
+                <pre id="state-transcribed">[The transcribed text will be displayed here]</pre>
+            </div>
+
+            <hr>
+
+            Debug output:
+            <textarea id="output" rows="20"></textarea>
+
+            <br>
+
+            <b>Troubleshooting</b>
+
+            <br><br>
+
+            The page does some heavy computations, so make sure:
+
+            <ul>
+                <li>To use a modern web browser (e.g. Chrome, Firefox)</li>
+                <li>To use a fast desktop or laptop computer (i.e. not a mobile phone)</li>
+                <li>Your browser supports WASM <a href="https://webassembly.org/roadmap/">Fixed-width SIMD</a></li>
+            </ul>
+
+            <div class="cell-version">
+                <span>
+                    |
+                    Build time: <span class="nav-link">@GIT_DATE@</span> |
+                    Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
+                    Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
+                    <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/stream.wasm">Source Code</a> |
+                </span>
+            </div>
+        </div>
+
+        <script type="text/javascript" src="helpers.js"></script>
+        <script type='text/javascript'>
+            // web audio context
+            var context = null;
+
+            // audio data
+            var audio = null;
+            var audio0 = null;
+
+            // the stream instance
+            var instance = null;
+
+            // model name
+            var model_whisper = null;
+
+            var Module = {
+                print: printTextarea,
+                printErr: printTextarea,
+                setStatus: function(text) {
+                    printTextarea('js: ' + text);
+                },
+                monitorRunDependencies: function(left) {
+                },
+                preRun: function() {
+                    printTextarea('js: Preparing ...');
+                },
+                postRun: function() {
+                    printTextarea('js: Initialized successfully!');
+                }
+            };
+
+            //
+            // fetch models
+            //
+
+            let dbVersion = 1
+            let dbName    = 'whisper.ggerganov.com';
+            let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
+
+            function storeFS(fname, buf) {
+                // write to WASM file using FS_createDataFile
+                // if the file exists, delete it
+                try {
+                    Module.FS_unlink(fname);
+                } catch (e) {
+                    // ignore
+                }
+
+                Module.FS_createDataFile("/", fname, buf, true, true);
+
+                printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
+
+                document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
+
+                if (model_whisper != null) {
+                    document.getElementById('start').disabled = false;
+                    document.getElementById('stop' ).disabled = true;
+                }
+            }
+
+            function loadWhisper(model) {
+                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',
+                };
+
+                let sizes = {
+                    'tiny.en': 75,
+                    'base.en': 142,
+                };
+
+                let url     = urls[model];
+                let dst     = 'whisper.bin';
+                let size_mb = sizes[model];
+
+                model_whisper = model;
+
+                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
+
+                cbProgress = function(p) {
+                    let el = document.getElementById('fetch-whisper-progress');
+                    el.innerHTML = Math.round(100*p) + '%';
+                };
+
+                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 = '';
+                };
+
+                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
+            }
+
+            //
+            // microphone
+            //
+
+            const kSampleRate = 16000;
+            const kRestartRecording_s = 120;
+            const kIntervalAudio_ms = 5000; // pass the recorded audio to the C++ instance at this rate
+
+            var mediaRecorder = null;
+            var doRecording = false;
+            var startTime = 0;
+
+            window.AudioContext = window.AudioContext || window.webkitAudioContext;
+            window.OfflineAudioContext = window.OfflineAudioContext || window.webkitOfflineAudioContext;
+
+            function stopRecording() {
+                Module.set_status("paused");
+                doRecording = false;
+                audio0 = null;
+                audio = null;
+                context = null;
+            }
+
+            function startRecording() {
+                if (!context) {
+                    context = new AudioContext({
+                        sampleRate: kSampleRate,
+                        channelCount: 1,
+                        echoCancellation: false,
+                        autoGainControl:  true,
+                        noiseSuppression: true,
+                    });
+                }
+
+                Module.set_status("");
+
+                document.getElementById('start').disabled = true;
+                document.getElementById('stop').disabled = false;
+
+                doRecording = true;
+                startTime = Date.now();
+
+                var chunks = [];
+                var stream = null;
+
+                navigator.mediaDevices.getUserMedia({audio: true, video: false})
+                    .then(function(s) {
+                        stream = s;
+                        mediaRecorder = new MediaRecorder(stream);
+                        mediaRecorder.ondataavailable = function(e) {
+                            chunks.push(e.data);
+
+                            var blob = new Blob(chunks, { 'type' : 'audio/ogg; codecs=opus' });
+                            var reader = new FileReader();
+
+                            reader.onload = function(event) {
+                                var buf = new Uint8Array(reader.result);
+
+                                if (!context) {
+                                    return;
+                                }
+                                context.decodeAudioData(buf.buffer, function(audioBuffer) {
+                                    var offlineContext = new OfflineAudioContext(audioBuffer.numberOfChannels, audioBuffer.length, audioBuffer.sampleRate);
+                                    var source = offlineContext.createBufferSource();
+                                    source.buffer = audioBuffer;
+                                    source.connect(offlineContext.destination);
+                                    source.start(0);
+
+                                    offlineContext.startRendering().then(function(renderedBuffer) {
+                                        audio = renderedBuffer.getChannelData(0);
+
+                                        //printTextarea('js: audio recorded, size: ' + audio.length + ', old size: ' + (audio0 == null ? 0 : audio0.length));
+
+                                        var audioAll = new Float32Array(audio0 == null ? audio.length : audio0.length + audio.length);
+                                        if (audio0 != null) {
+                                            audioAll.set(audio0, 0);
+                                        }
+                                        audioAll.set(audio, audio0 == null ? 0 : audio0.length);
+
+                                        if (instance) {
+                                            Module.set_audio(instance, audioAll);
+                                        }
+                                    });
+                                }, function(e) {
+                                    audio = null;
+                                });
+                            }
+
+                            reader.readAsArrayBuffer(blob);
+                        };
+
+                        mediaRecorder.onstop = function(e) {
+                            if (doRecording) {
+                                setTimeout(function() {
+                                    startRecording();
+                                });
+                            }
+                        };
+
+                        mediaRecorder.start(kIntervalAudio_ms);
+                    })
+                    .catch(function(err) {
+                        printTextarea('js: error getting audio stream: ' + err);
+                    });
+
+                var interval = setInterval(function() {
+                    if (!doRecording) {
+                        clearInterval(interval);
+                        mediaRecorder.stop();
+                        stream.getTracks().forEach(function(track) {
+                            track.stop();
+                        });
+
+                        document.getElementById('start').disabled = false;
+                        document.getElementById('stop').disabled  = true;
+
+                        mediaRecorder = null;
+                    }
+
+                    // if audio length is more than kRestartRecording_s seconds, restart recording
+                    if (audio != null && audio.length > kSampleRate*kRestartRecording_s) {
+                        if (doRecording) {
+                            //printTextarea('js: restarting recording');
+
+                            clearInterval(interval);
+                            audio0 = audio;
+                            audio = null;
+                            mediaRecorder.stop();
+                            stream.getTracks().forEach(function(track) {
+                                track.stop();
+                            });
+                        }
+                    }
+                }, 100);
+            }
+
+            //
+            // main
+            //
+
+            var nLines = 0;
+            var intervalUpdate = null;
+            var transcribedAll = '';
+
+            function onStart() {
+                if (!instance) {
+                    instance = Module.init('whisper.bin');
+
+                    if (instance) {
+                        printTextarea("js: whisper initialized, instance: " + instance);
+                    }
+                }
+
+                if (!instance) {
+                    printTextarea("js: failed to initialize whisper");
+                    return;
+                }
+
+                startRecording();
+
+                intervalUpdate = setInterval(function() {
+                    var transcribed = Module.get_transcribed();
+
+                    if (transcribed != null && transcribed.length > 1) {
+                        transcribedAll += transcribed + '<br>';
+                        nLines++;
+
+                        // if more than 10 lines, remove the first line
+                        if (nLines > 10) {
+                            var i = transcribedAll.indexOf('<br>');
+                            if (i > 0) {
+                                transcribedAll = transcribedAll.substring(i + 4);
+                                nLines--;
+                            }
+                        }
+                    }
+
+                    document.getElementById('state-status').innerHTML = Module.get_status();
+                    document.getElementById('state-transcribed').innerHTML = transcribedAll;
+                }, 100);
+            }
+
+            function onStop() {
+                stopRecording();
+            }
+
+        </script>
+        <script type="text/javascript" src="stream.js"></script>
+    </body>
+</html>

examples/stream/CMakeLists.txt

@@ -2,6 +2,8 @@ if (WHISPER_SUPPORT_SDL2)
     # stream
     set(TARGET stream)
     add_executable(${TARGET} stream.cpp)
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/stream/README.md

@@ -10,6 +10,23 @@ More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/i
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
 
+## Sliding window mode with VAD
+
+Setting the `--step` argument to `0` enables the sliding window mode:
+
+```java
+ ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
+```
+
+In this mode, the tool will transcribe only after some speech activity is detected. A very
+basic VAD detector is used, but in theory a more sophisticated approach can be added. The
+`-vth` argument determines the VAD threshold - higher values will make it detect silence more often.
+It's best to tune it to the specific use case, but a value around `0.6` should be OK in general.
+When silence is detected, it will transcribe the last `--length` milliseconds of audio and output
+a transcription block that is suitable for parsing.
+
+## Building
+
 The `stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
@@ -21,3 +38,7 @@ brew install sdl2
 
 make stream
 ```
+
+## Web version
+
+This tool can also run in the browser: [examples/stream.wasm](/examples/stream.wasm)

examples/stream/stream.cpp

@@ -1,17 +1,12 @@
 // Real-time speech recognition of input from a microphone
 //
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
+//
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 
-// third-party utilities
-// use your favorite implementations
-#define DR_WAV_IMPLEMENTATION
-#include "dr_wav.h"
-
-#include <SDL.h>
-#include <SDL_audio.h>
-
 #include <cassert>
 #include <cstdio>
 #include <string>
@@ -35,21 +30,27 @@ std::string to_timestamp(int64_t t) {
 
 // command-line parameters
 struct whisper_params {
-    int32_t seed       = -1; // RNG seed, not used currently
     int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
     int32_t step_ms    = 3000;
     int32_t length_ms  = 10000;
+    int32_t keep_ms    = 200;
     int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
 
-    bool verbose              = false;
-    bool translate            = false;
-    bool no_context           = true;
-    bool print_special_tokens = false;
-    bool no_timestamps        = true;
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool no_fallback   = false;
+    bool print_special = false;
+    bool no_context    = true;
+    bool no_timestamps = false;
 
     std::string language  = "en";
     std::string model     = "models/ggml-base.en.bin";
-    std::string fname_out = "";
+    std::string fname_out;
 };
 
 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@@ -58,41 +59,28 @@ 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 == "-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 == "--step") {
-            params.step_ms = std::stoi(argv[++i]);
-        } else if (arg == "--length") {
-            params.length_ms = std::stoi(argv[++i]);
-        } else if (arg == "-c" || arg == "--capture") {
-            params.capture_id = std::stoi(argv[++i]);
-        } else if (arg == "-v" || arg == "--verbose") {
-            params.verbose = true;
-        } else if (arg == "--translate") {
-            params.translate = true;
-        } else if (arg == "-kc" || arg == "--keep-context") {
-            params.no_context = false;
-        } else if (arg == "-l" || arg == "--language") {
-            params.language = argv[++i];
-            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);
-            }
-        } else if (arg == "-ps" || arg == "--print_special") {
-            params.print_special_tokens = true;
-        } else if (arg == "-nt" || arg == "--no_timestamps") {
-            params.no_timestamps = true;
-        } else if (arg == "-m" || arg == "--model") {
-            params.model = argv[++i];
-        } else if (arg == "-f" || arg == "--file") {
-            params.fname_out = argv[++i];
-        } else if (arg == "-h" || arg == "--help") {
+        if (arg == "-h" || arg == "--help") {
             whisper_print_usage(argc, argv, params);
             exit(0);
-        } else {
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (                 arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
+        else if (                 arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
+        else if (                 arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
             exit(0);
@@ -102,96 +90,32 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     return true;
 }
 
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params) {
+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           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, "            --step N         audio step size in milliseconds (default: %d)\n", params.step_ms);
-    fprintf(stderr, "            --length N       audio length in milliseconds (default: %d)\n", params.length_ms);
-    fprintf(stderr, "  -c ID,    --capture ID     capture device ID (default: -1)\n");
-    fprintf(stderr, "  -v,       --verbose        verbose output\n");
-    fprintf(stderr, "            --translate      translate from source language to english\n");
-    fprintf(stderr, "  -kc,      --keep-context   keep text context from earlier audio (default: false)\n");
-    fprintf(stderr, "  -ps,      --print_special  print special tokens\n");
-    fprintf(stderr, "  -nt,      --no_timestamps  do not print timestamps\n");
-    fprintf(stderr, "  -l LANG,  --language LANG  spoken language (default: %s)\n", params.language.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME    model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME     text output file name (default: no output to file)\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, "            --step N        [%-7d] audio step size in milliseconds\n",                params.step_ms);
+    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                   params.length_ms);
+    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",         params.keep_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, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",              params.no_context ? "false" : "true");
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
+    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-SDL_AudioDeviceID g_dev_id_in = 0;
-
-bool audio_sdl_init(const int capture_id) {
-    if (g_dev_id_in) {
-        fprintf(stderr, "%s: already initialized\n", __func__);
-        return false;
-    }
-
-    if (g_dev_id_in == 0) {
-        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 (1);
-        }
-
-        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));
-            }
-        }
-    }
-
-    if (g_dev_id_in == 0) {
-        SDL_AudioSpec capture_spec_requested;
-        SDL_AudioSpec capture_spec_obtained;
-
-        SDL_zero(capture_spec_requested);
-        SDL_zero(capture_spec_obtained);
-
-        capture_spec_requested.freq     = WHISPER_SAMPLE_RATE;
-        capture_spec_requested.format   = AUDIO_F32;
-        capture_spec_requested.channels = 1;
-        capture_spec_requested.samples  = 1024;
-
-        if (capture_id >= 0) {
-            fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-            g_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__);
-            g_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-        }
-        if (!g_dev_id_in) {
-            fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-            g_dev_id_in = 0;
-        } else {
-            fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, g_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);
-        }
-    }
-
-
-    return true;
-}
-
-///////////////////////////
-
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -199,29 +123,47 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.seed < 0) {
-        params.seed = time(NULL);
-    }
+    params.keep_ms   = std::min(params.keep_ms,   params.step_ms);
+    params.length_ms = std::max(params.length_ms, params.step_ms);
+
+    const int n_samples_step = (1e-3*params.step_ms  )*WHISPER_SAMPLE_RATE;
+    const int n_samples_len  = (1e-3*params.length_ms)*WHISPER_SAMPLE_RATE;
+    const int n_samples_keep = (1e-3*params.keep_ms  )*WHISPER_SAMPLE_RATE;
+    const int n_samples_30s  = (1e-3*30000.0         )*WHISPER_SAMPLE_RATE;
+
+    const bool use_vad = n_samples_step <= 0; // sliding window mode uses VAD
+
+    const int n_new_line = !use_vad ? std::max(1, params.length_ms / params.step_ms - 1) : 1; // number of steps to print new line
+
+    params.no_timestamps  = !use_vad;
+    params.no_context    |= use_vad;
+    params.max_tokens     = 0;
 
     // init audio
 
-    if (!audio_sdl_init(params.capture_id)) {
-        fprintf(stderr, "%s: audio_sdl_init() failed!\n", __func__);
+    audio_async audio(params.length_ms);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
         return 1;
     }
 
+    audio.resume();
+
     // whisper init
 
-    struct whisper_context * ctx = whisper_init(params.model.c_str());
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
 
-    const int n_samples = (params.step_ms/1000.0)*WHISPER_SAMPLE_RATE;
-    const int n_samples_len = (params.length_ms/1000.0)*WHISPER_SAMPLE_RATE;
-    const int n_samples_30s = 30*WHISPER_SAMPLE_RATE;
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
 
-    std::vector<float> pcmf32(n_samples_30s, 0.0f);
+    std::vector<float> pcmf32    (n_samples_30s, 0.0f);
     std::vector<float> pcmf32_old;
+    std::vector<float> pcmf32_new(n_samples_30s, 0.0f);
 
-    const int n_new_line = params.length_ms / params.step_ms - 1;
+    std::vector<whisper_token> prompt_tokens;
 
     // print some info about the processing
     {
@@ -233,23 +175,28 @@ 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 %d samples (step = %.1f sec / len = %.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+        fprintf(stderr, "%s: processing %d samples (step = %.1f sec / len = %.1f sec / keep = %.1f sec), %d threads, lang = %s, task = %s, timestamps = %d ...\n",
                 __func__,
-                n_samples,
-                float(n_samples)/WHISPER_SAMPLE_RATE,
-                float(n_samples_len)/WHISPER_SAMPLE_RATE,
+                n_samples_step,
+                float(n_samples_step)/WHISPER_SAMPLE_RATE,
+                float(n_samples_len )/WHISPER_SAMPLE_RATE,
+                float(n_samples_keep)/WHISPER_SAMPLE_RATE,
                 params.n_threads,
                 params.language.c_str(),
                 params.translate ? "translate" : "transcribe",
                 params.no_timestamps ? 0 : 1);
 
-        fprintf(stderr, "%s: n_new_line = %d\n", __func__, n_new_line);
+        if (!use_vad) {
+            fprintf(stderr, "%s: n_new_line = %d, no_context = %d\n", __func__, n_new_line, params.no_context);
+        } else {
+            fprintf(stderr, "%s: using VAD, will transcribe on speech activity\n", __func__);
+        }
+
         fprintf(stderr, "\n");
     }
 
-    SDL_PauseAudioDevice(g_dev_id_in, 0);
-
     int n_iter = 0;
+
     bool is_running = true;
 
     std::ofstream fout;
@@ -264,67 +211,100 @@ int main(int argc, char ** argv) {
     printf("[Start speaking]");
     fflush(stdout);
 
+          auto t_last  = std::chrono::high_resolution_clock::now();
+    const auto t_start = t_last;
+
     // main audio loop
     while (is_running) {
-        // process SDL events:
-        SDL_Event event;
-        while (SDL_PollEvent(&event)) {
-            switch (event.type) {
-                case SDL_QUIT:
-                    {
-                        is_running = false;
-                    } break;
-                default:
-                    break;
-            }
-        }
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
 
         if (!is_running) {
             break;
         }
 
         // process new audio
-        if (n_iter > 0 && SDL_GetQueuedAudioSize(g_dev_id_in) > 2*n_samples*sizeof(float)) {
-            fprintf(stderr, "\n\n%s: WARNING: cannot process audio fast enough, dropping audio ...\n\n", __func__);
-            SDL_ClearQueuedAudio(g_dev_id_in);
+
+        if (!use_vad) {
+            while (true) {
+                audio.get(params.step_ms, pcmf32_new);
+
+                if ((int) pcmf32_new.size() > 2*n_samples_step) {
+                    fprintf(stderr, "\n\n%s: WARNING: cannot process audio fast enough, dropping audio ...\n\n", __func__);
+                    audio.clear();
+                    continue;
+                }
+
+                if ((int) pcmf32_new.size() >= n_samples_step) {
+                    audio.clear();
+                    break;
+                }
+
+                std::this_thread::sleep_for(std::chrono::milliseconds(1));
+            }
+
+            const int n_samples_new = pcmf32_new.size();
+
+            // take up to params.length_ms audio from previous iteration
+            const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_keep + n_samples_len - n_samples_new));
+
+            //printf("processing: take = %d, new = %d, old = %d\n", n_samples_take, n_samples_new, (int) pcmf32_old.size());
+
+            pcmf32.resize(n_samples_new + n_samples_take);
+
+            for (int i = 0; i < n_samples_take; i++) {
+                pcmf32[i] = pcmf32_old[pcmf32_old.size() - n_samples_take + i];
+            }
+
+            memcpy(pcmf32.data() + n_samples_take, pcmf32_new.data(), n_samples_new*sizeof(float));
+
+            pcmf32_old = pcmf32;
+        } else {
+            const auto t_now  = std::chrono::high_resolution_clock::now();
+            const auto t_diff = std::chrono::duration_cast<std::chrono::milliseconds>(t_now - t_last).count();
+
+            if (t_diff < 2000) {
+                std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+                continue;
+            }
+
+            audio.get(2000, pcmf32_new);
+
+            if (::vad_simple(pcmf32_new, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, false)) {
+                audio.get(params.length_ms, pcmf32);
+            } else {
+                std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+                continue;
+            }
+
+            t_last = t_now;
         }
 
-        while (SDL_GetQueuedAudioSize(g_dev_id_in) < n_samples*sizeof(float)) {
-            SDL_Delay(1);
-        }
-
-        const int n_samples_new = SDL_GetQueuedAudioSize(g_dev_id_in)/sizeof(float);
-
-        // take one second from previous iteration
-        //const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_30s/30 - n_samples_new));
-
-        // take up to params.length_ms audio from previous iteration
-        const int n_samples_take = std::min((int) pcmf32_old.size(), std::max(0, n_samples_len - n_samples_new));
-
-        //printf("processing: take = %d, new = %d, old = %d\n", n_samples_take, n_samples_new, (int) pcmf32_old.size());
-
-        pcmf32.resize(n_samples_new + n_samples_take);
-
-        for (int i = 0; i < n_samples_take; i++) {
-            pcmf32[i] = pcmf32_old[pcmf32_old.size() - n_samples_take + i];
-        }
-
-        SDL_DequeueAudio(g_dev_id_in, pcmf32.data() + n_samples_take, n_samples_new*sizeof(float));
-
-        pcmf32_old = pcmf32;
-
         // run the inference
         {
             whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
 
-            wparams.print_progress       = false;
-            wparams.print_special_tokens = params.print_special_tokens;
-            wparams.print_realtime       = false;
-            wparams.print_timestamps     = !params.no_timestamps;
-            wparams.translate            = params.translate;
-            wparams.no_context           = params.no_context;
-            wparams.language             = params.language.c_str();
-            wparams.n_threads            = params.n_threads;
+            wparams.print_progress   = false;
+            wparams.print_special    = params.print_special;
+            wparams.print_realtime   = false;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.translate        = params.translate;
+            wparams.single_segment   = !use_vad;
+            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;
+
+            // disable temperature fallback
+            //wparams.temperature_inc  = -1.0f;
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
+
+            wparams.prompt_tokens    = params.no_context ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens  = params.no_context ? 0       : prompt_tokens.size();
 
             if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
@@ -333,12 +313,21 @@ int main(int argc, char ** argv) {
 
             // print result;
             {
-                printf("\33[2K\r");
+                if (!use_vad) {
+                    printf("\33[2K\r");
 
-                // print long empty line to clear the previous line
-                printf("%s", std::string(100, ' ').c_str());
+                    // print long empty line to clear the previous line
+                    printf("%s", std::string(100, ' ').c_str());
 
-                printf("\33[2K\r");
+                    printf("\33[2K\r");
+                } else {
+                    const int64_t t1 = (t_last - t_start).count()/1000000;
+                    const int64_t t0 = std::max(0.0, t1 - pcmf32.size()*1000.0/WHISPER_SAMPLE_RATE);
+
+                    printf("\n");
+                    printf("### Transcription %d START | t0 = %d ms | t1 = %d ms\n", n_iter, (int) t0, (int) t1);
+                    printf("\n");
+                }
 
                 const int n_segments = whisper_full_n_segments(ctx);
                 for (int i = 0; i < n_segments; ++i) {
@@ -366,18 +355,39 @@ int main(int argc, char ** argv) {
                 if (params.fname_out.length() > 0) {
                     fout << std::endl;
                 }
+
+                if (use_vad){
+                    printf("\n");
+                    printf("### Transcription %d END\n", n_iter);
+                }
             }
 
             ++n_iter;
 
-            if ((n_iter % n_new_line) == 0) {
+            if (!use_vad && (n_iter % n_new_line) == 0) {
                 printf("\n");
 
-                pcmf32_old.clear();
+                // keep part of the audio for next iteration to try to mitigate word boundary issues
+                pcmf32_old = std::vector<float>(pcmf32.end() - n_samples_keep, pcmf32.end());
+
+                // Add tokens of the last full length segment as the prompt
+                if (!params.no_context) {
+                    prompt_tokens.clear();
+
+                    const int n_segments = whisper_full_n_segments(ctx);
+                    for (int i = 0; i < n_segments; ++i) {
+                        const int token_count = whisper_full_n_tokens(ctx, i);
+                        for (int j = 0; j < token_count; ++j) {
+                            prompt_tokens.push_back(whisper_full_get_token_id(ctx, i, j));
+                        }
+                    }
+                }
             }
         }
     }
 
+    audio.pause();
+
     whisper_print_timings(ctx);
     whisper_free(ctx);
 

examples/talk-llama/.gitignore

@@ -0,0 +1 @@
+audio.mp3

examples/talk-llama/CMakeLists.txt

@@ -0,0 +1,16 @@
+if (WHISPER_SUPPORT_SDL2)
+    # talk-llama
+    set(TARGET talk-llama)
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+endif ()

examples/talk-llama/README.md

@@ -0,0 +1,36 @@
+# talk-llama
+
+Talk with an LLaMA AI in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
+
+## Building
+
+The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk-llama" executable
+make talk-llama
+
+# Run it
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
+- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
+By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
+
+## Discussion
+
+If you have any feedback, please let "us" know in the following discussion: https://github.com/ggerganov/whisper.cpp/discussions/672?converting=1

examples/talk-llama/eleven-labs.py

@@ -0,0 +1,23 @@
+import sys
+import importlib.util
+
+api_key = "" #Write your https://beta.elevenlabs.io api key here
+if not api_key:
+    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk-llama/eleven-labs.py")
+    sys.exit()
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import ElevenLabs
+eleven = ElevenLabs(api_key)
+
+# Get a Voice object, by name or UUID
+voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = voice.generate(str(sys.argv[2:]))
+
+# Save the TTS to a file
+audio.save("audio") 

examples/talk-llama/llama.h

@@ -0,0 +1,173 @@
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#define LLAMA_FILE_VERSION 1
+#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
+#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id;  // token id
+
+        float p;     // probability of the token
+        float plog;  // log probability of the token
+
+    } llama_token_data;
+
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
+
+    struct llama_context_params {
+        int n_ctx;   // text context
+        int n_parts; // -1 for default
+        int seed;    // RNG seed, 0 for random
+
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // TODO: not great API - very likely to change
+    // Returns 0 on success
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+                   int   itype);
+
+    // Returns the KV cache that will contain the context for the
+    // ongoing prediction with the model.
+    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
+
+    // Returns the size of the KV cache
+    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+
+    // Sets the KV cache containing the current context for the model
+    LLAMA_API void llama_set_kv_cache(
+            struct llama_context * ctx,
+                   const uint8_t * kv_cache,
+                          size_t   n_size,
+                             int   n_token_count);
+
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * 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 a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_eos();
+
+    // TODO: improve the last_n_tokens interface ?
+    LLAMA_API llama_token llama_sample_top_p_top_k(
+       struct llama_context * ctx,
+          const llama_token * last_n_tokens_data,
+                        int   last_n_tokens_size,
+                        int   top_k,
+                      float   top_p,
+                      float   temp,
+                      float   repeat_penalty);
+
+    // Performance information
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // LLAMA_H