diarization : try conv and self-attention embeddings

This seems to be causing hallucinations in the end of the audio, e.g.:

"Thank you for listening"
"Amen"
..

.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

@@ -1,267 +1,267 @@
 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
+  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
+    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
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Add msbuild to PATH
-              uses: microsoft/setup-msbuild@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" >> $env:GITHUB_ENV
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ 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: 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: 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/${{ matrix.build }}
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$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 }}
+      - 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
+  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
+    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
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
 
-            - name: Add msbuild to PATH
-              uses: microsoft/setup-msbuild@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 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" >> $env:GITHUB_ENV
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ 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: 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: 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 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/${{ matrix.build }}
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$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 }}
+      - 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
+  emscripten:
+    runs-on: ubuntu-latest
 
-        strategy:
-            matrix:
-                build: [Release]
+    strategy:
+      matrix:
+        build: [Release]
 
-        steps:
-            - name: Clone
-              uses: actions/checkout@v1
+    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: 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: Configure
+        run: echo "tmp"
 
-            - name: Build
-              run: |
-                pushd emsdk-master
-                source ./emsdk_env.sh
-                popd
-                emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-                make
+      - name: Build
+        run: |
+          pushd emsdk-master
+          source ./emsdk_env.sh
+          popd
+          emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          make

.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

@@ -1,4 +1,5 @@
 *.o
+*.a
 .cache/
 .vs/
 .vscode/
@@ -8,6 +9,7 @@ build/
 build-em/
 build-debug/
 build-release/
+build-static/
 build-sanitize-addr/
 build-sanitize-thread/
 
@@ -17,7 +19,9 @@ build-sanitize-thread/
 /talk
 /bench
 
+arm_neon.h
 sync.sh
+libwhisper.a
 libwhisper.so
 compile_commands.json
 

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.0.4)
+project(whisper.cpp VERSION 1.2.0)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -226,10 +226,13 @@ target_compile_definitions(${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
     )
 
 #
@@ -242,7 +245,7 @@ add_subdirectory(bindings)
 # programs, examples and tests
 #
 
-if (WHISPER_BUILD_TESTS)
+if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     enable_testing()
     add_subdirectory(tests)
 endif ()

Makefile

@@ -115,11 +115,15 @@ endif
 ifeq ($(UNAME_M),amd64)
 	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
-ifeq ($(UNAME_M),ppc64le)
+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
@@ -133,8 +137,8 @@ ifdef WHISPER_OPENBLAS
 	LDFLAGS += -lopenblas
 endif
 ifdef WHISPER_GPROF
-	CFLAGS  += -pg
-	CXXFLAGS  += -pg
+	CFLAGS   += -pg
+	CXXFLAGS += -pg
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 endif
@@ -193,18 +197,21 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-main: examples/main/main.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o whisper.o -o main $(LDFLAGS)
+SRC_COMMON = examples/common.cpp
+SRC_COMMON_SDL = examples/common-sdl.cpp
+
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o -o main $(LDFLAGS)
 	./main -h
 
-stream: examples/stream/stream.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
 
-command: examples/command/command.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
 
-talk: examples/talk/talk.cpp  examples/talk/gpt-2.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
 
 bench: examples/bench/bench.cpp ggml.o whisper.o
 	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)

README.md

@@ -4,7 +4,7 @@
 [![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/)
 
-[Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.2.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -13,7 +13,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - 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)
@@ -71,7 +71,7 @@ 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
 ```
 
 ---
@@ -89,27 +89,38 @@ 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          [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
-  -wt N,    --word-thold N  [0.01   ] word timestamp probability threshold
-  -su,      --speed-up      [false  ] speed up audio by x2 (reduced accuracy)
-  -tr,      --translate     [false  ] translate from source language to english
-  -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
-  -ps,      --print-special [false  ] print special tokens
-  -pc,      --print-colors  [false  ] print colors
-  -nt,      --no-timestamps [true   ] do not print timestamps
-  -l LANG,  --language LANG [en     ] spoken language
-  -m FNAME, --model FNAME   [models/ggml-base.en.bin] model path
-  -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 ...
@@ -128,7 +139,8 @@ 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
@@ -141,13 +153,14 @@ 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  =  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: 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 ctx     =  140.60 MB
 whisper_model_load: model size    =  140.54 MB
 
-system_info: n_threads = 4 / 10 | AVX = 0 | 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 ...
 
@@ -155,12 +168,13 @@ 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`.
@@ -203,26 +217,16 @@ 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 | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
+| 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` |
 
 ## 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.
+- No GPU support (yet)
 
 ## Another example
 
@@ -235,7 +239,8 @@ 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
@@ -248,55 +253,60 @@ 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>
 
@@ -307,6 +317,7 @@ The [stream](examples/stream) tool samples the audio every half a second and run
 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
 ```
 
@@ -321,14 +332,14 @@ to highlight words with high or low confidence:
 
 ## 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 ...
 
@@ -352,7 +363,7 @@ 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 ...
 
@@ -453,7 +464,11 @@ in [models](models).
 - [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
 - [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
 - [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)
+- [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)
 - [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
 
 ## Examples

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

@@ -2,6 +2,12 @@ package main
 
 import (
 	"flag"
+	"fmt"
+	"strings"
+	"time"
+
+	// Packages
+	whisper "github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
 )
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -42,6 +48,26 @@ 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"
 }
@@ -50,12 +76,81 @@ 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", "", "Language")
+	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

@@ -35,8 +35,7 @@ func main() {
 
 	// Process files
 	for _, filename := range flags.Args() {
-		fmt.Println("Processing", filename)
-		if err := Process(model, filename, flags.GetLanguage(), flags.IsSpeedup(), flags.IsTokens()); err != nil {
+		if err := Process(model, filename, flags); err != nil {
 			fmt.Fprintln(os.Stderr, err)
 			continue
 		}

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

@@ -11,7 +11,7 @@ import (
 	wav "github.com/go-audio/wav"
 )
 
-func Process(model whisper.Model, path string, lang string, speedup, tokens bool) error {
+func Process(model whisper.Model, path string, flags *Flags) error {
 	var data []float32
 
 	// Create processing context
@@ -20,14 +20,22 @@ func Process(model whisper.Model, path string, lang string, speedup, tokens bool
 		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
+	// Decode the WAV file - load the full buffer
 	dec := wav.NewDecoder(fh)
 	if buf, err := dec.FullPCMBuffer(); err != nil {
 		return err
@@ -39,42 +47,86 @@ func Process(model whisper.Model, path string, lang string, speedup, tokens bool
 		data = buf.AsFloat32Buffer().Data
 	}
 
-	// Set the parameters
+	// Segment callback when -tokens is specified
 	var cb whisper.SegmentCallback
-	if lang != "" {
-		if err := context.SetLanguage(lang); err != nil {
-			return err
-		}
-	}
-	if speedup {
-		context.SetSpeedup(true)
-	}
-	if tokens {
+	if flags.IsTokens() {
 		cb = func(segment whisper.Segment) {
-			fmt.Printf("%02d [%6s->%6s] ", segment.Num, segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond))
+			fmt.Fprintf(flags.Output(), "%02d [%6s->%6s] ", segment.Num, segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond))
 			for _, token := range segment.Tokens {
-				fmt.Printf("%q ", token.Text)
+				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.Println("")
+			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 {
-			break
+			return nil
 		} else if err != nil {
 			return err
 		}
-		fmt.Printf("[%6s->%6s] %s\n", segment.Start.Truncate(time.Millisecond), segment.End.Truncate(time.Millisecond), segment.Text)
+		fmt.Fprintln(w, n)
+		fmt.Fprintln(w, srtTimestamp(segment.Start), " --> ", srtTimestamp(segment.End))
+		fmt.Fprintln(w, segment.Text)
+		fmt.Fprintln(w, "")
+		n++
 	}
-
-	// Return success
-	return nil
+}
+
+// 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/params.go

@@ -47,7 +47,12 @@ 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
@@ -57,6 +62,7 @@ func (p *Params) SetLanguage(lang int) error {
 	return nil
 }
 
+// Get language id
 func (p *Params) Language() int {
 	if p.language == nil {
 		return -1
@@ -64,18 +70,46 @@ func (p *Params) Language() int {
 	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)
+}
+
+// Set max tokens per segment (0 = no limit)
+func (p *Params) SetMaxTokensPerSegment(n int) {
+	p.max_tokens = C.int(n)
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS
 

bindings/go/pkg/whisper/consts.go

@@ -11,10 +11,11 @@ import (
 // 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")
+	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")
 )
 
 ///////////////////////////////////////////////////////////////////////////////

bindings/go/pkg/whisper/context.go

@@ -1,7 +1,9 @@
 package whisper
 
 import (
+	"fmt"
 	"io"
+	"runtime"
 	"strings"
 	"time"
 
@@ -24,7 +26,7 @@ var _ Context = (*context)(nil)
 ///////////////////////////////////////////////////////////////////////////////
 // LIFECYCLE
 
-func NewContext(model *model, params whisper.Params) (Context, error) {
+func newContext(model *model, params whisper.Params) (Context, error) {
 	context := new(context)
 	context.model = model
 	context.params = params
@@ -41,7 +43,13 @@ func (context *context) SetLanguage(lang string) error {
 	if context.model.ctx == nil {
 		return ErrInternalAppError
 	}
-	if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
+	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
@@ -50,16 +58,94 @@ func (context *context) SetLanguage(lang string) error {
 	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 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 {
@@ -119,6 +205,65 @@ func (context *context) NextSegment() (Segment, error) {
 	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
 

bindings/go/pkg/whisper/interface.go

@@ -20,15 +20,28 @@ type Model interface {
 	// 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.
+	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
-	SetSpeedup(bool)          // Set speedup flag
+
+	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
+	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
@@ -38,6 +51,21 @@ type Context interface {
 	// 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.

bindings/go/pkg/whisper/model.go

@@ -23,7 +23,7 @@ var _ Model = (*model)(nil)
 ///////////////////////////////////////////////////////////////////////////////
 // LIFECYCLE
 
-func New(path string) (*model, error) {
+func New(path string) (Model, error) {
 	model := new(model)
 	if _, err := os.Stat(path); err != nil {
 		return nil, err
@@ -64,6 +64,11 @@ func (model *model) String() string {
 ///////////////////////////////////////////////////////////////////////////////
 // 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())
@@ -91,5 +96,5 @@ func (model *model) NewContext() (Context, error) {
 	params.SetThreads(runtime.NumCPU())
 
 	// Return new context
-	return NewContext(model, params)
+	return newContext(model, params)
 }

bindings/go/whisper.go

@@ -91,7 +91,7 @@ var (
 func Whisper_init(path string) *Context {
 	cPath := C.CString(path)
 	defer C.free(unsafe.Pointer(cPath))
-	if ctx := C.whisper_init(cPath); ctx != nil {
+	if ctx := C.whisper_init_from_file(cPath); ctx != nil {
 		return (*Context)(ctx)
 	} else {
 		return nil
@@ -147,16 +147,6 @@ func (ctx *Context) Whisper_decode(tokens []Token, past, threads int) error {
 	}
 }
 
-// whisper_sample_best() returns the token with the highest probability
-func (ctx *Context) Whisper_sample_best() TokenData {
-	return TokenData(C.whisper_sample_best((*C.struct_whisper_context)(ctx)))
-}
-
-// whisper_sample_timestamp() returns the most probable timestamp token
-func (ctx *Context) Whisper_sample_timestamp(is_initial bool) TokenData {
-	return TokenData(C.whisper_sample_timestamp((*C.struct_whisper_context)(ctx), C.bool(is_initial)))
-}
-
 // 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) {

bindings/javascript/emscripten.cpp

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

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.0.4",
+  "version": "1.2.0",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/.gitignore

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

bindings/ruby/ext/extconf.rb

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

bindings/ruby/ext/ruby_whisper.cpp

@@ -0,0 +1,426 @@
+#include <ruby.h>
+#include "ruby_whisper.h"
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BOOL_PARAMS_SETTER(self, prop, value) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (value == Qfalse || value == Qnil) { \
+    rwp->params.prop = false; \
+  } else { \
+    rwp->params.prop = true; \
+  } \
+  return value; \
+
+#define BOOL_PARAMS_GETTER(self,  prop) \
+  ruby_whisper_params *rwp; \
+  Data_Get_Struct(self, ruby_whisper_params, rwp); \
+  if (rwp->params.prop) { \
+    return Qtrue; \
+  } else { \
+    return Qfalse; \
+  }
+
+VALUE mWhisper;
+VALUE cContext;
+VALUE cParams;
+
+static void ruby_whisper_free(ruby_whisper *rw) {
+  if (rw->context) {
+    whisper_free(rw->context);
+    rw->context = NULL;
+  }
+}
+static void ruby_whisper_params_free(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_mark(ruby_whisper *rw) {
+  // call rb_gc_mark on any ruby references in rw
+}
+
+void rb_whisper_free(ruby_whisper *rw) {
+  ruby_whisper_free(rw);
+  free(rw);
+}
+
+void rb_whisper_params_mark(ruby_whisper_params *rwp) {
+}
+
+void rb_whisper_params_free(ruby_whisper_params *rwp) {
+  ruby_whisper_params_free(rwp);
+  free(rwp);
+}
+
+static VALUE ruby_whisper_allocate(VALUE klass) {
+  ruby_whisper *rw;
+  rw = ALLOC(ruby_whisper);
+  rw->context = NULL;
+  return Data_Wrap_Struct(klass, rb_whisper_mark, rb_whisper_free, rw);
+}
+
+static VALUE ruby_whisper_params_allocate(VALUE klass) {
+  ruby_whisper_params *rwp;
+  rwp = ALLOC(ruby_whisper_params);
+  rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  return Data_Wrap_Struct(klass, rb_whisper_params_mark, rb_whisper_params_free, rwp);
+}
+
+static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  VALUE whisper_model_file_path;
+
+  // TODO: we can support init from buffer here too maybe another ruby object to expose
+  rb_scan_args(argc, argv, "01", &whisper_model_file_path);
+  Data_Get_Struct(self, ruby_whisper, rw);
+
+  if (!rb_respond_to(whisper_model_file_path, rb_intern("to_s"))) {
+    rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
+  }
+  rw->context = whisper_init_from_file(StringValueCStr(whisper_model_file_path));
+  if (rw->context == nullptr) {
+    rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
+  }
+  return self;
+}
+
+/*
+ * transcribe a single file
+ * can emit to a block results
+ *
+ **/
+static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  VALUE wave_file_path, blk, params;
+
+  rb_scan_args(argc, argv, "02&", &wave_file_path, &params, &blk);
+  Data_Get_Struct(self, ruby_whisper, rw);
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+
+  if (!rb_respond_to(wave_file_path, rb_intern("to_s"))) {
+    rb_raise(rb_eRuntimeError, "Expected file path to wave file");
+  }
+
+  std::string fname_inp = StringValueCStr(wave_file_path);
+
+  std::vector<float> pcmf32; // mono-channel F32 PCM
+  std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+  // WAV input - this is directly from main.cpp example
+  {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname_inp == "-") {
+      {
+        uint8_t buf[1024];
+        while (true) {
+          const size_t n = fread(buf, 1, sizeof(buf), stdin);
+          if (n == 0) {
+            break;
+          }
+          wav_data.insert(wav_data.end(), buf, buf + n);
+        }
+      }
+
+      if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open WAV file from stdin\n");
+        return self;
+      }
+
+      fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    } else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
+      fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+      fprintf(stderr, "WAV file '%s' must be mono or stereo\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (rwp->diarize && wav.channels != 2 && rwp->params.print_timestamps == false) {
+      fprintf(stderr, "WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", fname_inp.c_str());
+      return self;
+    }
+
+    if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
+      fprintf(stderr, "WAV file '%s' must be %i kHz\n", fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
+      return self;
+    }
+
+    if (wav.bitsPerSample != 16) {
+      fprintf(stderr, "WAV file '%s' must be 16-bit\n", fname_inp.c_str());
+      return self;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[i])/32768.0f;
+      }
+    } else {
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+      }
+    }
+
+    if (rwp->diarize) {
+      // convert to stereo, float
+      pcmf32s.resize(2);
+
+      pcmf32s[0].resize(n);
+      pcmf32s[1].resize(n);
+      for (uint64_t i = 0; i < n; i++) {
+        pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+        pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+      }
+    }
+  }
+  {
+    static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, 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

examples/CMakeLists.txt

@@ -14,6 +14,37 @@ 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})
@@ -24,6 +55,8 @@ if (EMSCRIPTEN)
     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)

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 +1,15 @@
+const path = require('path');
+const { whisper } = require(path.join(__dirname, '../../../build/Release/whisper-addon'));
+
+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", () => {
+        expect(whisper(whisperParamsMock).length).toBeGreaterThan(0);
+    });
+});

examples/addon.node/addon.cpp

@@ -0,0 +1,342 @@
+#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, 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;
+    }
+
+    // initial prompt
+    std::vector<whisper_token> prompt_tokens;
+
+    if (!params.prompt.empty()) {
+        prompt_tokens.resize(1024);
+        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+        fprintf(stderr, "\n");
+        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
+        fprintf(stderr, "initial tokens: [ ");
+        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
+            fprintf(stderr, "%d ", prompt_tokens[i]);
+        }
+        fprintf(stderr, "]\n");
+    }
+
+    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
+        const auto fname_inp = params.fname_inp[f];
+        const auto fname_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.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+            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*/, 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;
+}
+
+Napi::Object whisper(const Napi::CallbackInfo& info) {
+    Napi::Env env = info.Env();
+    if (info.Length() <= 0 || !info[0].IsObject()) {
+        Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
+    }
+    whisper_params params;
+    std::vector<std::vector<std::string>> result;
+
+    Napi::Object whisper_params = info[0].As<Napi::Object>();
+    std::string language = whisper_params.Get("language").As<Napi::String>();
+    std::string model = whisper_params.Get("model").As<Napi::String>();
+    std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+
+    params.language = language;
+    params.model = model;
+    params.fname_inp.emplace_back(input);
+
+    // run model
+    run(params, result);
+
+    fprintf(stderr, "RESULT:\n");
+    for (auto sentence:result) {
+        fprintf(stderr, "t0: %s, t1: %s, content: %s \n",
+                sentence[0].c_str(), sentence[1].c_str(), sentence[2].c_str());
+    }
+
+    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;
+    }
+
+    return res;
+}
+
+
+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 +1,27 @@
+const path = require('path');
+const { whisper } = require(path.join(__dirname, '../../build/Release/whisper-addon'));
+
+const whisperParams = {
+    language: 'en',
+    model: path.join(__dirname, '../../models/ggml-base.en.bin'),
+    fname_inp: '',
+};
+
+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);
+console.log(whisper(whisperParams));

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/emscripten.cpp

@@ -28,6 +28,11 @@ void bench_main(size_t index) {
         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;
@@ -52,7 +57,7 @@ EMSCRIPTEN_BINDINGS(bench) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                 if (g_contexts[i] != nullptr) {
                     if (g_worker.joinable()) {
                         g_worker.join();

examples/bench/bench.cpp

@@ -7,6 +7,7 @@
 // command-line parameters
 struct whisper_params {
     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";
 };
@@ -23,6 +24,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         }
         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);
@@ -41,19 +43,17 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     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");
@@ -92,3 +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

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

examples/command.wasm/emscripten.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "common.h"
 #include "whisper.h"
 
 #include <emscripten.h>
@@ -27,24 +28,6 @@ std::string g_transcribed   = "";
 
 std::vector<float> g_pcmf32;
 
-static std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-static void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
 // compute similarity between two strings using Levenshtein distance
 static float similarity(const std::string & s0, const std::string & s1) {
     const size_t len0 = s0.size() + 1;
@@ -75,44 +58,6 @@ void command_set_status(const std::string & status) {
     g_status = status;
 }
 
-bool command_vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (size_t i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string command_transcribe(whisper_context * ctx, const whisper_full_params & wparams, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -155,7 +100,7 @@ void command_get_audio(int ms, int sample_rate, std::vector<float> & audio) {
     const int64_t n_samples = (ms * sample_rate) / 1000;
 
     int64_t n_take = 0;
-    if (g_pcmf32.size() < n_samples) {
+    if (n_samples > (int) g_pcmf32.size()) {
         n_take = g_pcmf32.size();
     } else {
         n_take = n_samples;
@@ -187,7 +132,6 @@ void command_main(size_t index) {
 
     printf("command: using %d threads\n", wparams.n_threads);
 
-    bool is_running   = true;
     bool have_prompt  = false;
     bool ask_prompt   = true;
     bool print_energy = false;
@@ -233,7 +177,7 @@ void command_main(size_t index) {
         {
             command_get_audio(vad_ms, WHISPER_SAMPLE_RATE, pcmf32_cur);
 
-            if (command_vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, vad_thold, freq_thold, print_energy)) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
                 command_set_status("Speech detected! Processing ...");
 
@@ -324,7 +268,7 @@ EMSCRIPTEN_BINDINGS(command) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/command/CMakeLists.txt

@@ -5,6 +5,5 @@ if (WHISPER_SUPPORT_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/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/helpers.js

@@ -8,7 +8,7 @@ function convertTypedArray(src, type) {
 
 var printTextarea = (function() {
     var element = document.getElementById('output');
-    if (element) element.alue = ''; // clear browser cache
+    if (element) element.value = ''; // clear browser cache
     return function(text) {
         if (arguments.length > 1) text = Array.prototype.slice.call(arguments).join(' ');
         console.log(text);
@@ -88,11 +88,15 @@ async function fetchRemote(url, cbProgress, cbPrint) {
 // - 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) {
-    // 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');
-    });
+    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);

examples/livestream.sh

@@ -100,7 +100,7 @@ while [ $running -eq 1 ]; do
         err=$(cat /tmp/whisper-live.err | wc -l)
     done
 
-    ./main -t 8 -m ./models/ggml-base.en.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
+    ./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

examples/main/CMakeLists.txt

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

examples/main/README.md

@@ -9,25 +9,35 @@ It can be used as a reference for using the `whisper.cpp` library in other proje
 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
-  -wt N,    --word-thold N  [0.01   ] word timestamp probability threshold
-  -su,      --speed-up      [false  ] speed up audio by x2 (reduced accuracy)
-  -tr,      --translate     [false  ] translate from source language to english
-  -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
-  -ps,      --print-special [false  ] print special tokens
-  -pc,      --print-colors  [false  ] print colors
-  -nt,      --no-timestamps [true   ] do not print timestamps
-  -l LANG,  --language LANG [en     ] spoken language
-  -m FNAME, --model FNAME   [models/ggml-base.en.bin] model path
-  -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
 ```

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>
@@ -53,18 +50,24 @@ void replace_all(std::string & s, const std::string & search, const std::string
 // command-line parameters
 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 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      =  5;
+    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 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;
@@ -80,6 +83,7 @@ struct whisper_params {
     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);
@@ -88,6 +92,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
 
+        if (arg == "-"){
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
         if (arg[0] != '-') {
             params.fname_inp.push_back(arg);
             continue;
@@ -104,15 +113,22 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         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 == "-ocsv" || arg == "--output-csv")     { params.output_csv     = 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; }
@@ -136,31 +152,38 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help           [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N      [%-7d] number of threads to use during computation\n",    params.n_threads);
-    fprintf(stderr, "  -p N,     --processors N   [%-7d] number of processors to use during computation\n", params.n_processors);
-    fprintf(stderr, "  -ot N,    --offset-t N     [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
-    fprintf(stderr, "  -on N,    --offset-n N     [%-7d] segment index offset\n",                           params.offset_n);
-    fprintf(stderr, "  -d  N,    --duration N     [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
-    fprintf(stderr, "  -mc N,    --max-context N  [%-7d] maximum number of text context tokens to store\n", params.max_context);
-    fprintf(stderr, "  -ml N,    --max-len N      [%-7d] maximum segment length in characters\n",           params.max_len);
-    fprintf(stderr, "  -wt N,    --word-thold N   [%-7.2f] word timestamp probability threshold\n",         params.word_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, "  -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, "  -owts,    --output-words   [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
-    fprintf(stderr, "  -ocsv,    --output-csv     [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pc,      --print-colors   [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
-    fprintf(stderr, "  -pp,      --print-progress [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
-    fprintf(stderr, "  -nt,      --no-timestamps  [%-7s] do not print timestamps\n",                        params.no_timestamps ? "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, "  -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, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "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");
 }
 
@@ -235,7 +258,7 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
                 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()))));
+                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");
             }
@@ -331,20 +354,16 @@ bool output_csv(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);
-	if (text[0] == ' ')
-	  text = text + sizeof(char); //whisper_full_get_segment_text() returns a string with leading space, point to the next character.
         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";
+
+        //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;
 }
 
-
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
@@ -478,7 +497,7 @@ int main(int argc, char ** argv) {
 
     // 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");
@@ -503,90 +522,14 @@ 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];
 
-        std::vector<float> pcmf32; // mono-channel F32 PCM
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
         std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
 
-        // WAV input
-        {
-            drwav wav;
-            std::vector<uint8_t> wav_data; // used for pipe input from stdin
-
-            if (fname_inp == "-") {
-                {
-                    uint8_t buf[1024];
-                    while (true)
-                    {
-                        const size_t n = fread(buf, 1, sizeof(buf), stdin);
-                        if (n == 0) {
-                            break;
-                        }
-                        wav_data.insert(wav_data.end(), buf, buf + n);
-                    }
-                }
-
-                if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-                    fprintf(stderr, "error: failed to open WAV file from stdin\n");
-                    return 4;
-                }
-
-                fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
-            }
-            else if (drwav_init_file(&wav, fname_inp.c_str(), nullptr) == false) {
-                fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname_inp.c_str());
-                return 5;
-            }
-
-            if (wav.channels != 1 && wav.channels != 2) {
-                fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (params.diarize && wav.channels != 2 && params.no_timestamps == false) {
-                fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization and timestamps have to be enabled\n", argv[0], fname_inp.c_str());
-                return 6;
-            }
-
-            if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-                fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", argv[0], fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
-                return 8;
-            }
-
-            if (wav.bitsPerSample != 16) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", argv[0], fname_inp.c_str());
-                return 9;
-            }
-
-            const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
-
-            std::vector<int16_t> pcm16;
-            pcm16.resize(n*wav.channels);
-            drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
-            drwav_uninit(&wav);
-
-            // convert to mono, float
-            pcmf32.resize(n);
-            if (wav.channels == 1) {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[i])/32768.0f;
-                }
-            } else {
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
-                }
-            }
-
-            if (params.diarize) {
-                // convert to stereo, float
-                pcmf32s.resize(2);
-
-                pcmf32s[0].resize(n);
-                pcmf32s[1].resize(n);
-                for (uint64_t i = 0; i < n; i++) {
-                    pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
-                    pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
-                }
-            }
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
         }
 
         // print system information
@@ -620,6 +563,8 @@ int main(int argc, char ** argv) {
         {
             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;
@@ -634,11 +579,19 @@ int main(int argc, char ** argv) {
             wparams.token_timestamps = params.output_wts || params.max_len > 0;
             wparams.thold_pt         = params.word_thold;
             wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+            wparams.split_on_word    = params.split_on_word;
 
             wparams.speed_up         = params.speed_up;
 
-            wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+            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 };
 
@@ -665,6 +618,8 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                 return 10;
             }
+
+            whisper_full_cluster_segments(ctx);
         }
 
         // output stuff
@@ -673,34 +628,33 @@ 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
+            // output to CSV file
             if (params.output_csv) {
-                const auto fname_csv = fname_inp + ".csv";
+                const auto fname_csv = fname_out + ".csv";
                 output_csv(ctx, fname_csv.c_str());
             }
-
         }
     }
 

examples/stream.wasm/emscripten.cpp

@@ -49,6 +49,9 @@ void stream_main(size_t index) {
     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);
@@ -129,7 +132,7 @@ EMSCRIPTEN_BINDINGS(stream) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/stream/CMakeLists.txt

@@ -5,6 +5,5 @@ if (WHISPER_SUPPORT_SDL2)
 
     include(DefaultTargetOptions)
 
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/stream/stream.cpp

@@ -3,19 +3,16 @@
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
 //
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
-#include <atomic>
 #include <cassert>
 #include <cstdio>
 #include <string>
 #include <thread>
 #include <vector>
 #include <fstream>
-#include <mutex>
 
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
@@ -116,306 +113,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    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;
-};
-
-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));
-        }
-    }
-}
-
-///////////////////////////
-
-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;
-}
-
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -423,20 +120,21 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    params.keep_ms = std::min(params.keep_ms, params.step_ms); // cannot be more than step_ms
+    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 = (params.step_ms  *1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_len  = (params.length_ms*1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_keep = (params.keep_ms  *1e-3)*WHISPER_SAMPLE_RATE;
-    const int n_samples_30s  = (30000           *1e-3)*WHISPER_SAMPLE_RATE;
+    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 ? params.length_ms / params.step_ms - 1 : 1; // number of steps to print new line
+    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;
+    params.no_timestamps  = !use_vad;
+    params.no_context    |= use_vad;
+    params.max_tokens     = 0;
 
     // init audio
 
@@ -456,10 +154,10 @@ int main(int argc, char ** argv) {
         exit(0);
     }
 
-    struct whisper_context * ctx = whisper_init(params.model.c_str());
+    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_old(n_samples_30s, 0.0f);
+    std::vector<float> pcmf32_old;
     std::vector<float> pcmf32_new(n_samples_30s, 0.0f);
 
     std::vector<whisper_token> prompt_tokens;
@@ -486,7 +184,7 @@ int main(int argc, char ** argv) {
                 params.no_timestamps ? 0 : 1);
 
         if (!use_vad) {
-            fprintf(stderr, "%s: n_new_line = %d\n", __func__, n_new_line);
+            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__);
         }
@@ -516,23 +214,7 @@ int main(int argc, char ** argv) {
     // main audio loop
     while (is_running) {
         // handle Ctrl + C
-        {
-            SDL_Event event;
-            while (SDL_PollEvent(&event)) {
-                switch (event.type) {
-                    case SDL_QUIT:
-                        {
-                            is_running = false;
-                        } break;
-                    default:
-                        break;
-                }
-            }
-
-            if (!is_running) {
-                break;
-            }
-        }
+        is_running = sdl_poll_events();
 
         if (!is_running) {
             break;
@@ -555,7 +237,7 @@ int main(int argc, char ** argv) {
                     break;
                 }
 
-                SDL_Delay(1);
+                std::this_thread::sleep_for(std::chrono::milliseconds(1));
             }
 
             const int n_samples_new = pcmf32_new.size();
@@ -586,7 +268,7 @@ int main(int argc, char ** argv) {
 
             audio.get(2000, pcmf32_new);
 
-            if (vad_simple(pcmf32_new, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, false)) {
+            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));
@@ -615,6 +297,9 @@ int main(int argc, char ** argv) {
             wparams.audio_ctx        = params.audio_ctx;
             wparams.speed_up         = params.speed_up;
 
+            // disable temperature fallback
+            wparams.temperature_inc  = -1.0f;
+
             wparams.prompt_tokens    = params.no_context ? nullptr : prompt_tokens.data();
             wparams.prompt_n_tokens  = params.no_context ? 0       : prompt_tokens.size();
 

examples/talk.wasm/emscripten.cpp

@@ -271,7 +271,7 @@ EMSCRIPTEN_BINDINGS(talk) {
     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());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                 if (g_contexts[i] != nullptr) {
                     g_running = true;
                     if (g_worker.joinable()) {

examples/talk/CMakeLists.txt

@@ -7,7 +7,7 @@ if (WHISPER_SUPPORT_SDL2)
 
     # TODO: this is temporary
     #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk.cpp gpt-2.cpp ../../ggml.c ../../whisper.cpp)
+    add_executable(${TARGET} talk.cpp gpt-2.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
 
     include(DefaultTargetOptions)
 

examples/talk/talk.cpp

@@ -1,16 +1,14 @@
 // Talk with AI
 //
 
+#include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 #include "gpt-2.h"
 
-#include <SDL.h>
-#include <SDL_audio.h>
-
 #include <cassert>
 #include <cstdio>
 #include <fstream>
-#include <mutex>
 #include <regex>
 #include <string>
 #include <thread>
@@ -105,320 +103,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
 }
 
-//
-// SDL Audio capture
-//
-
-class audio_async {
-public:
-    audio_async(int len_ms);
-    ~audio_async();
-
-    bool init(int capture_id, int sample_rate);
-
-    // start capturing audio via the provided SDL callback
-    // keep last len_ms seconds of audio in a circular buffer
-    bool resume();
-    bool pause();
-    bool clear();
-
-    // callback to be called by SDL
-    void callback(uint8_t * stream, int len);
-
-    // get audio data from the circular buffer
-    void get(int ms, std::vector<float> & audio);
-
-private:
-    SDL_AudioDeviceID m_dev_id_in = 0;
-
-    int m_len_ms = 0;
-    int m_sample_rate = 0;
-
-    bool       m_running = false;
-    std::mutex m_mutex;
-
-    std::vector<float> m_audio;
-    std::vector<float> m_audio_new;
-    size_t             m_audio_pos = 0;
-    size_t             m_audio_len = 0;
-};
-
-audio_async::audio_async(int len_ms) {
-    m_len_ms = len_ms;
-}
-
-audio_async::~audio_async() {
-    if (m_dev_id_in) {
-        SDL_CloseAudioDevice(m_dev_id_in);
-    }
-}
-
-bool audio_async::init(int capture_id, int sample_rate) {
-    SDL_LogSetPriority(SDL_LOG_CATEGORY_APPLICATION, SDL_LOG_PRIORITY_INFO);
-
-    if (SDL_Init(SDL_INIT_AUDIO) < 0) {
-        SDL_LogError(SDL_LOG_CATEGORY_APPLICATION, "Couldn't initialize SDL: %s\n", SDL_GetError());
-        return false;
-    }
-
-    SDL_SetHintWithPriority(SDL_HINT_AUDIO_RESAMPLING_MODE, "medium", SDL_HINT_OVERRIDE);
-
-    {
-        int nDevices = SDL_GetNumAudioDevices(SDL_TRUE);
-        fprintf(stderr, "%s: found %d capture devices:\n", __func__, nDevices);
-        for (int i = 0; i < nDevices; i++) {
-            fprintf(stderr, "%s:    - Capture device #%d: '%s'\n", __func__, i, SDL_GetAudioDeviceName(i, SDL_TRUE));
-        }
-    }
-
-    SDL_AudioSpec capture_spec_requested;
-    SDL_AudioSpec capture_spec_obtained;
-
-    SDL_zero(capture_spec_requested);
-    SDL_zero(capture_spec_obtained);
-
-    capture_spec_requested.freq     = sample_rate;
-    capture_spec_requested.format   = AUDIO_F32;
-    capture_spec_requested.channels = 1;
-    capture_spec_requested.samples  = 1024;
-    capture_spec_requested.callback = [](void * userdata, uint8_t * stream, int len) {
-        audio_async * audio = (audio_async *) userdata;
-        audio->callback(stream, len);
-    };
-    capture_spec_requested.userdata = this;
-
-    if (capture_id >= 0) {
-        fprintf(stderr, "%s: attempt to open capture device %d : '%s' ...\n", __func__, capture_id, SDL_GetAudioDeviceName(capture_id, SDL_TRUE));
-        m_dev_id_in = SDL_OpenAudioDevice(SDL_GetAudioDeviceName(capture_id, SDL_TRUE), SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    } else {
-        fprintf(stderr, "%s: attempt to open default capture device ...\n", __func__);
-        m_dev_id_in = SDL_OpenAudioDevice(nullptr, SDL_TRUE, &capture_spec_requested, &capture_spec_obtained, 0);
-    }
-
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: couldn't open an audio device for capture: %s!\n", __func__, SDL_GetError());
-        m_dev_id_in = 0;
-
-        return false;
-    } else {
-        fprintf(stderr, "%s: obtained spec for input device (SDL Id = %d):\n", __func__, m_dev_id_in);
-        fprintf(stderr, "%s:     - sample rate:       %d\n",                   __func__, capture_spec_obtained.freq);
-        fprintf(stderr, "%s:     - format:            %d (required: %d)\n",    __func__, capture_spec_obtained.format,
-                capture_spec_requested.format);
-        fprintf(stderr, "%s:     - channels:          %d (required: %d)\n",    __func__, capture_spec_obtained.channels,
-                capture_spec_requested.channels);
-        fprintf(stderr, "%s:     - samples per frame: %d\n",                   __func__, capture_spec_obtained.samples);
-        fprintf(stderr, "\n");
-    }
-
-    m_sample_rate = capture_spec_obtained.freq;
-
-    m_audio.resize((m_sample_rate*m_len_ms)/1000);
-
-    return true;
-}
-
-bool audio_async::resume() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to resume!\n", __func__);
-        return false;
-    }
-
-    if (m_running) {
-        fprintf(stderr, "%s: already running!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 0);
-
-    m_running = true;
-
-    return true;
-}
-
-bool audio_async::pause() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to pause!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: already paused!\n", __func__);
-        return false;
-    }
-
-    SDL_PauseAudioDevice(m_dev_id_in, 1);
-
-    m_running = false;
-
-    return true;
-}
-
-bool audio_async::clear() {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to clear!\n", __func__);
-        return false;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return false;
-    }
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        m_audio_pos = 0;
-        m_audio_len = 0;
-    }
-
-    return true;
-}
-
-// callback to be called by SDL
-void audio_async::callback(uint8_t * stream, int len) {
-    if (!m_running) {
-        return;
-    }
-
-    const size_t n_samples = len / sizeof(float);
-
-    m_audio_new.resize(n_samples);
-    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));
-
-    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (m_audio_pos + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - m_audio_pos;
-
-            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = m_audio.size();
-        } else {
-            memcpy(&m_audio[m_audio_pos], stream, n_samples * sizeof(float));
-
-            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
-            m_audio_len = std::min(m_audio_len + n_samples, m_audio.size());
-        }
-    }
-}
-
-void audio_async::get(int ms, std::vector<float> & result) {
-    if (!m_dev_id_in) {
-        fprintf(stderr, "%s: no audio device to get audio from!\n", __func__);
-        return;
-    }
-
-    if (!m_running) {
-        fprintf(stderr, "%s: not running!\n", __func__);
-        return;
-    }
-
-    result.clear();
-
-    {
-        std::lock_guard<std::mutex> lock(m_mutex);
-
-        if (ms <= 0) {
-            ms = m_len_ms;
-        }
-
-        size_t n_samples = (m_sample_rate * ms) / 1000;
-        if (n_samples > m_audio_len) {
-            n_samples = m_audio_len;
-        }
-
-        result.resize(n_samples);
-
-        int s0 = m_audio_pos - n_samples;
-        if (s0 < 0) {
-            s0 += m_audio.size();
-        }
-
-        if (s0 + n_samples > m_audio.size()) {
-            const size_t n0 = m_audio.size() - s0;
-
-            memcpy(result.data(), &m_audio[s0], n0 * sizeof(float));
-            memcpy(&result[n0], &m_audio[0], (n_samples - n0) * sizeof(float));
-        } else {
-            memcpy(result.data(), &m_audio[s0], n_samples * sizeof(float));
-        }
-    }
-}
-
-///////////////////////////
-
-std::string trim(const std::string & s) {
-    std::regex e("^\\s+|\\s+$");
-    return std::regex_replace(s, e, "");
-}
-
-std::string replace(const std::string & s, const std::string & from, const std::string & to) {
-    std::string result = s;
-    size_t pos = 0;
-    while ((pos = result.find(from, pos)) != std::string::npos) {
-        result.replace(pos, from.length(), to);
-        pos += to.length();
-    }
-    return result;
-}
-
-void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
-    const float rc = 1.0f / (2.0f * M_PI * cutoff);
-    const float dt = 1.0f / sample_rate;
-    const float alpha = dt / (rc + dt);
-
-    float y = data[0];
-
-    for (size_t i = 1; i < data.size(); i++) {
-        y = alpha * (y + data[i] - data[i - 1]);
-        data[i] = y;
-    }
-}
-
-bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
-    const int n_samples      = pcmf32.size();
-    const int n_samples_last = (sample_rate * last_ms) / 1000;
-
-    if (n_samples_last >= n_samples) {
-        // not enough samples - assume no speech
-        return false;
-    }
-
-    if (freq_thold > 0.0f) {
-        high_pass_filter(pcmf32, freq_thold, sample_rate);
-    }
-
-    float energy_all  = 0.0f;
-    float energy_last = 0.0f;
-
-    for (int i = 0; i < n_samples; i++) {
-        energy_all += fabsf(pcmf32[i]);
-
-        if (i >= n_samples - n_samples_last) {
-            energy_last += fabsf(pcmf32[i]);
-        }
-    }
-
-    energy_all  /= n_samples;
-    energy_last /= n_samples_last;
-
-    if (verbose) {
-        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
-    }
-
-    if (energy_last > vad_thold*energy_all) {
-        return false;
-    }
-
-    return true;
-}
-
 std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
     const auto t_start = std::chrono::high_resolution_clock::now();
 
@@ -498,7 +182,7 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    struct whisper_context * ctx_wsp = whisper_init(params.model_wsp.c_str());
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
 
     // gpt init
 
@@ -557,22 +241,10 @@ int main(int argc, char ** argv) {
     // main loop
     while (is_running) {
         // handle Ctrl + C
-        {
-            SDL_Event event;
-            while (SDL_PollEvent(&event)) {
-                switch (event.type) {
-                    case SDL_QUIT:
-                        {
-                            is_running = false;
-                        } break;
-                    default:
-                        break;
-                }
-            }
+        is_running = sdl_poll_events();
 
-            if (!is_running) {
-                break;
-            }
+        if (!is_running) {
+            break;
         }
 
         // delay
@@ -583,7 +255,7 @@ int main(int argc, char ** argv) {
         {
             audio.get(2000, pcmf32_cur);
 
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
                 fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
 
                 audio.get(params.voice_ms, pcmf32_cur);

examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt

@@ -64,16 +64,21 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
     private suspend fun copyAssets() = withContext(Dispatchers.IO) {
         modelsPath.mkdirs()
         samplesPath.mkdirs()
-        application.copyData("models", modelsPath, ::printMessage)
+        //application.copyData("models", modelsPath, ::printMessage)
         application.copyData("samples", samplesPath, ::printMessage)
         printMessage("All data copied to working directory.\n")
     }
 
     private suspend fun loadBaseModel() = withContext(Dispatchers.IO) {
         printMessage("Loading model...\n")
-        val firstModel = modelsPath.listFiles()!!.first()
-        whisperContext = WhisperContext.createContext(firstModel.absolutePath)
-        printMessage("Loaded model ${firstModel.name}.\n")
+        val models = application.assets.list("models/")
+        if (models != null) {
+            whisperContext = WhisperContext.createContextFromAsset(application.assets, "models/" + models[0])
+            printMessage("Loaded model ${models[0]}.\n")
+        }
+
+        //val firstModel = modelsPath.listFiles()!!.first()
+        //whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
     }
 
     fun transcribeSample() = viewModelScope.launch {

examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt

@@ -1,9 +1,11 @@
 package com.whispercppdemo.whisper
 
+import android.content.res.AssetManager
 import android.os.Build
 import android.util.Log
 import kotlinx.coroutines.*
 import java.io.File
+import java.io.InputStream
 import java.util.concurrent.Executors
 
 private const val LOG_TAG = "LibWhisper"
@@ -39,13 +41,31 @@ class WhisperContext private constructor(private var ptr: Long) {
     }
 
     companion object {
-        fun createContext(filePath: String): WhisperContext {
+        fun createContextFromFile(filePath: String): WhisperContext {
             val ptr = WhisperLib.initContext(filePath)
             if (ptr == 0L) {
                 throw java.lang.RuntimeException("Couldn't create context with path $filePath")
             }
             return WhisperContext(ptr)
         }
+
+        fun createContextFromInputStream(stream: InputStream): WhisperContext {
+            val ptr = WhisperLib.initContextFromInputStream(stream)
+
+            if (ptr == 0L) {
+                throw java.lang.RuntimeException("Couldn't create context from input stream")
+            }
+            return WhisperContext(ptr)
+        }
+
+        fun createContextFromAsset(assetManager: AssetManager, assetPath: String): WhisperContext {
+            val ptr = WhisperLib.initContextFromAsset(assetManager, assetPath)
+
+            if (ptr == 0L) {
+                throw java.lang.RuntimeException("Couldn't create context from asset $assetPath")
+            }
+            return WhisperContext(ptr)
+        }
     }
 }
 
@@ -76,6 +96,8 @@ private class WhisperLib {
         }
 
         // JNI methods
+        external fun initContextFromInputStream(inputStream: InputStream): Long
+        external fun initContextFromAsset(assetManager: AssetManager, assetPath: String): Long
         external fun initContext(modelPath: String): Long
         external fun freeContext(contextPtr: Long)
         external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)

examples/whisper.android/app/src/main/jni/whisper/Whisper.mk

@@ -1,5 +1,5 @@
 WHISPER_LIB_DIR := $(LOCAL_PATH)/../../../../../../../
-LOCAL_LDLIBS    := -llog
+LOCAL_LDLIBS    := -landroid -llog
 
 # Make the final output library smaller by only keeping the symbols referenced from the app.
 ifneq ($(APP_OPTIM),debug)

examples/whisper.android/app/src/main/jni/whisper/jni.c

@@ -1,13 +1,17 @@
 #include <jni.h>
+#include <android/asset_manager.h>
+#include <android/asset_manager_jni.h>
 #include <android/log.h>
 #include <stdlib.h>
 #include <sys/sysinfo.h>
+#include <string.h>
 #include "whisper.h"
 
 #define UNUSED(x) (void)(x)
 #define TAG "JNI"
 
 #define LOGI(...) __android_log_print(ANDROID_LOG_INFO,     TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN,     TAG, __VA_ARGS__)
 
 static inline int min(int a, int b) {
     return (a < b) ? a : b;
@@ -17,13 +21,132 @@ static inline int max(int a, int b) {
     return (a > b) ? a : b;
 }
 
+struct input_stream_context {
+    size_t offset;
+    JNIEnv * env;
+    jobject thiz;
+    jobject input_stream;
+
+    jmethodID mid_available;
+    jmethodID mid_read;
+};
+
+size_t inputStreamRead(void * ctx, void * output, size_t read_size) {
+    struct input_stream_context* is = (struct input_stream_context*)ctx;
+
+    jint avail_size = (*is->env)->CallIntMethod(is->env, is->input_stream, is->mid_available);
+    jint size_to_copy = read_size < avail_size ? (jint)read_size : avail_size;
+
+    jbyteArray byte_array = (*is->env)->NewByteArray(is->env, size_to_copy);
+
+    jint n_read = (*is->env)->CallIntMethod(is->env, is->input_stream, is->mid_read, byte_array, 0, size_to_copy);
+
+    if (size_to_copy != read_size || size_to_copy != n_read) {
+        LOGI("Insufficient Read: Req=%zu, ToCopy=%d, Available=%d", read_size, size_to_copy, n_read);
+    }
+
+    jbyte* byte_array_elements = (*is->env)->GetByteArrayElements(is->env, byte_array, NULL);
+    memcpy(output, byte_array_elements, size_to_copy);
+    (*is->env)->ReleaseByteArrayElements(is->env, byte_array, byte_array_elements, JNI_ABORT);
+
+    (*is->env)->DeleteLocalRef(is->env, byte_array);
+
+    is->offset += size_to_copy;
+
+    return size_to_copy;
+}
+bool inputStreamEof(void * ctx) {
+    struct input_stream_context* is = (struct input_stream_context*)ctx;
+
+    jint result = (*is->env)->CallIntMethod(is->env, is->input_stream, is->mid_available);
+    return result <= 0;
+}
+void inputStreamClose(void * ctx) {
+
+}
+
+JNIEXPORT jlong JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContextFromInputStream(
+        JNIEnv *env, jobject thiz, jobject input_stream) {
+    UNUSED(thiz);
+
+    struct whisper_context *context = NULL;
+    struct whisper_model_loader loader = {};
+    struct input_stream_context inp_ctx = {};
+
+    inp_ctx.offset = 0;
+    inp_ctx.env = env;
+    inp_ctx.thiz = thiz;
+    inp_ctx.input_stream = input_stream;
+
+    jclass cls = (*env)->GetObjectClass(env, input_stream);
+    inp_ctx.mid_available = (*env)->GetMethodID(env, cls, "available", "()I");
+    inp_ctx.mid_read = (*env)->GetMethodID(env, cls, "read", "([BII)I");
+
+    loader.context = &inp_ctx;
+    loader.read = inputStreamRead;
+    loader.eof = inputStreamEof;
+    loader.close = inputStreamClose;
+
+    loader.eof(loader.context);
+
+    context = whisper_init(&loader);
+    return (jlong) context;
+}
+
+static size_t asset_read(void *ctx, void *output, size_t read_size) {
+    return AAsset_read((AAsset *) ctx, output, read_size);
+}
+
+static bool asset_is_eof(void *ctx) {
+    return AAsset_getRemainingLength64((AAsset *) ctx) <= 0;
+}
+
+static void asset_close(void *ctx) {
+    AAsset_close((AAsset *) ctx);
+}
+
+static struct whisper_context *whisper_init_from_asset(
+        JNIEnv *env,
+        jobject assetManager,
+        const char *asset_path
+) {
+    LOGI("Loading model from asset '%s'\n", asset_path);
+    AAssetManager *asset_manager = AAssetManager_fromJava(env, assetManager);
+    AAsset *asset = AAssetManager_open(asset_manager, asset_path, AASSET_MODE_STREAMING);
+    if (!asset) {
+        LOGW("Failed to open '%s'\n", asset_path);
+        return NULL;
+    }
+
+    whisper_model_loader loader = {
+            .context = asset,
+            .read = &asset_read,
+            .eof = &asset_is_eof,
+            .close = &asset_close
+    };
+
+    return whisper_init(&loader);
+}
+
+JNIEXPORT jlong JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContextFromAsset(
+        JNIEnv *env, jobject thiz, jobject assetManager, jstring asset_path_str) {
+    UNUSED(thiz);
+    struct whisper_context *context = NULL;
+    const char *asset_path_chars = (*env)->GetStringUTFChars(env, asset_path_str, NULL);
+    context = whisper_init_from_asset(env, assetManager, asset_path_chars);
+    (*env)->ReleaseStringUTFChars(env, asset_path_str, asset_path_chars);
+    return (jlong) context;
+}
+
 JNIEXPORT jlong JNICALL
 Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_initContext(
         JNIEnv *env, jobject thiz, jstring model_path_str) {
     UNUSED(thiz);
     struct whisper_context *context = NULL;
     const char *model_path_chars = (*env)->GetStringUTFChars(env, model_path_str, NULL);
-    context = whisper_init(model_path_chars);
+    context = whisper_init_from_file(model_path_chars);
     (*env)->ReleaseStringUTFChars(env, model_path_str, model_path_chars);
     return (jlong) context;
 }

examples/whisper.android/local.properties

@@ -1,10 +0,0 @@
-## This file is automatically generated by Android Studio.
-# Do not modify this file -- YOUR CHANGES WILL BE ERASED!
-#
-# This file should *NOT* be checked into Version Control Systems,
-# as it contains information specific to your local configuration.
-#
-# Location of the SDK. This is only used by Gradle.
-# For customization when using a Version Control System, please read the
-# header note.
-sdk.dir=/Users/kevin/Library/Android/sdk

examples/whisper.objc/whisper.objc/ViewController.m

@@ -61,7 +61,7 @@ void AudioInputCallback(void * inUserData,
         NSLog(@"Loading model from %@", modelPath);
 
         // create ggml context
-        stateInp.ctx = whisper_init([modelPath UTF8String]);
+        stateInp.ctx = whisper_init_from_file([modelPath UTF8String]);
 
         // check if the model was loaded successfully
         if (stateInp.ctx == NULL) {

examples/whisper.swiftui/README.md

@@ -10,3 +10,5 @@ To use:
 5. Select the "release" build configuration under "Run", then deploy and run to your device.
 
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
+
+![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)

examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift

@@ -55,7 +55,7 @@ actor WhisperContext {
     }
     
     static func createContext(path: String) throws -> WhisperContext {
-        let context = whisper_init(path)
+        let context = whisper_init_from_file(path)
         if let context {
             return WhisperContext(context: context)
         } else {

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -35,10 +35,10 @@
 		0AAC5DA029539CD0003032C3 /* WhisperCppDemo.entitlements */ = {isa = PBXFileReference; lastKnownFileType = text.plist.entitlements; path = WhisperCppDemo.entitlements; sourceTree = "<group>"; };
 		0AAC5DA229539CD0003032C3 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		0AAC5DC629539EAF003032C3 /* WhisperCppDemo-Bridging-Header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "WhisperCppDemo-Bridging-Header.h"; sourceTree = "<group>"; };
-		0AAC5DC729539EB0003032C3 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../whisper.cpp; sourceTree = "<group>"; };
-		0AAC5DC829539EB0003032C3 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../whisper.h; sourceTree = "<group>"; };
-		0AAC5DC929539EB0003032C3 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml.c; sourceTree = "<group>"; };
-		0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml.h; sourceTree = "<group>"; };
+		0AAC5DC729539EB0003032C3 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; path = whisper.cpp; sourceTree = "<group>"; };
+		0AAC5DC829539EB0003032C3 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = whisper.h; sourceTree = "<group>"; };
+		0AAC5DC929539EB0003032C3 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = ggml.c; sourceTree = "<group>"; };
+		0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = ggml.h; sourceTree = "<group>"; };
 		0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
 		0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */
@@ -129,7 +129,8 @@
 				0AAC5DC729539EB0003032C3 /* whisper.cpp */,
 				0AAC5DC829539EB0003032C3 /* whisper.h */,
 			);
-			path = whisper.cpp;
+			name = whisper.cpp;
+			path = ../..;
 			sourceTree = "<group>";
 		};
 		0AAC5DCF2953A36C003032C3 /* whisper.cpp.swift */ = {

examples/whisper.wasm/CMakeLists.txt

@@ -32,8 +32,8 @@ 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 INITIAL_MEMORY=1500MB \
+    -s TOTAL_MEMORY=1500MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/whisper.wasm/emscripten.cpp

@@ -18,7 +18,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
 
         for (size_t i = 0; i < g_contexts.size(); ++i) {
             if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                 if (g_contexts[i] != nullptr) {
                     return i + 1;
                 } else {

examples/whisper.wasm/index-tmpl.html

@@ -46,10 +46,12 @@
 
             <div id="model">
                 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-tiny"    onclick="loadWhisper('tiny')">tiny (75 MB)</button>
-                <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
-                <button id="fetch-whisper-base"    onclick="loadWhisper('base')">base (142 MB)</button>
+                <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
+                <button id="fetch-whisper-tiny"     onclick="loadWhisper('tiny')">tiny (75 MB)</button>
+                <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <button id="fetch-whisper-base"     onclick="loadWhisper('base')">base (142 MB)</button>
+                <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
+                <button id="fetch-whisper-small"    onclick="loadWhisper('small')">small (466 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
@@ -60,8 +62,8 @@
             <!-- radio button to select between file upload or microphone -->
             <div id="input">
                 Input:
-                <input type="radio" id="file" name="input" value="file" checked="checked" onchange="changeInput('file')" /> File
-                <input type="radio" id="mic" name="input" value="mic" onchange="changeInput('mic')" /> Microphone
+                <input type="radio" id="file" name="input" value="file" checked="checked" onchange="changeInput('file')" /> <label for="file">File</label>
+                <input type="radio" id="mic" name="input" value="mic" onchange="changeInput('mic')" /> <label for="mic">Microphone</label>
             </div>
 
             <br>
@@ -284,27 +286,33 @@
                 }
                 reader.readAsArrayBuffer(file);
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
-                document.getElementById('fetch-whisper-tiny'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-base'   ).style.display = 'none';
-                document.getElementById('whisper-file'         ).style.display = 'none';
-                document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
+                document.getElementById('fetch-whisper-tiny-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-base'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-small'   ).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',
-                    'tiny':    'https://whisper.ggerganov.com/ggml-model-whisper-tiny.bin',
-                    'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
-                    'base':    'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
+                    'tiny.en':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
+                    'tiny':     'https://whisper.ggerganov.com/ggml-model-whisper-tiny.bin',
+                    'base.en':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'base':     'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
+                    'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
+                    'small':    'https://whisper.ggerganov.com/ggml-model-whisper-small.bin',
                 };
 
                 let sizes = {
-                    'tiny.en': 75,
-                    'tiny':    75,
-                    'base.en': 142,
-                    'base':    142,
+                    'tiny.en':  75,
+                    'tiny':     75,
+                    'base.en':  142,
+                    'base':     142,
+                    'small.en': 466,
+                    'small':    466,
                 };
 
                 let url     = urls[model];
@@ -313,12 +321,14 @@
 
                 model_whisper = model;
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
-                document.getElementById('fetch-whisper-tiny'   ).style.display = 'none';
-                document.getElementById('fetch-whisper-base'   ).style.display = 'none';
-                document.getElementById('whisper-file'         ).style.display = 'none';
-                document.getElementById('model-whisper-status' ).innerHTML = 'loading model: ' + model;
+                document.getElementById('fetch-whisper-tiny-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-base'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-small'   ).style.display = 'none';
+                document.getElementById('whisper-file'          ).style.display = 'none';
+                document.getElementById('model-whisper-status'  ).innerHTML = 'loading model: ' + model;
 
                 cbProgress = function(p) {
                     let el = document.getElementById('fetch-whisper-progress');
@@ -327,12 +337,14 @@
 
                 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('fetch-whisper-tiny'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('whisper-file'         ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-whisper-status' ); if (el) el.innerHTML = '';
+                    el = document.getElementById('fetch-whisper-tiny-en' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-tiny'    ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base'    ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small'   ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('whisper-file'          ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('model-whisper-status'  ); if (el) el.innerHTML = '';
                 };
 
                 loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);

examples/yt-wsp.sh

@@ -1,20 +1,10 @@
 #!/usr/bin/env bash
-
-# Small shell script to more easily automatically download and transcribe live stream VODs.
-# This uses YT-DLP, ffmpeg and the CPP version of Whisper: https://github.com/ggerganov/whisper.cpp
-# Use `./examples/yt-wsp.sh help` to print help info.
-#
-# Sample usage:
-#
-#   git clone https://github.com/ggerganov/whisper.cpp
-#   cd whisper.cpp
-#   make
-#   ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890
-#
+# shellcheck disable=2086
 
 # MIT License
 
 # Copyright (c) 2022 Daniils Petrovs
+# Copyright (c) 2023 Jennifer Capasso
 
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -34,114 +24,178 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
+# Small shell script to more easily automatically download and transcribe live stream VODs.
+# This uses YT-DLP, ffmpeg and the CPP version of Whisper: https://github.com/ggerganov/whisper.cpp
+# Use `./examples/yt-wsp.sh help` to print help info.
+#
+# Sample usage:
+#
+#   git clone https://github.com/ggerganov/whisper.cpp
+#   cd whisper.cpp
+#   make
+#   ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890
+#
+
 set -Eeuo pipefail
 
-# You can find how to download models in the OG repo: https://github.com/ggerganov/whisper.cpp/#usage
-MODEL_PATH="${MODEL_PATH:-models/ggml-base.en.bin}" # Set to a multilingual model if you want to translate from foreign lang to en
-WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-whisper}" # Where to find the whisper.cpp executable
-WHISPER_LANG="${WHISPER_LANG:-en}" # Set to desired lang to translate from
+# get script file location
+SCRIPT_PATH="$(realpath -e ${BASH_SOURCE[0]})";
+SCRIPT_DIR="${SCRIPT_PATH%/*}"
+
+################################################################################
+# Documentation on downloading models can be found in the whisper.cpp repo:
+# https://github.com/ggerganov/whisper.cpp/#usage
+#
+# note: unless a multilingual model is specified, WHISPER_LANG will be ignored
+# and the video will be transcribed as if the audio were in the English language
+################################################################################
+MODEL_PATH="${MODEL_PATH:-${SCRIPT_DIR}/../models/ggml-base.en.bin}"
+
+################################################################################
+# Where to find the whisper.cpp executable.  default to the examples directory
+# which holds this script in source control
+################################################################################
+WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-${SCRIPT_DIR}/../main}";
+
+# Set to desired language to be translated into english
+WHISPER_LANG="${WHISPER_LANG:-en}";
+
+# Default to 4 threads (this was most performant on my 2020 M1 MBP)
+WHISPER_THREAD_COUNT="${WHISPER_THREAD_COUNT:-4}";
 
 msg() {
     echo >&2 -e "${1-}"
 }
 
 cleanup() {
-    msg "Cleaning up..."
-    rm -rf "${temp_dir}" "vod-resampled.wav" "vod-resampled.wav.srt"
+    local -r clean_me="${1}";
+
+    if [ -d "${clean_me}" ]; then
+      msg "Cleaning up...";
+      rm -rf "${clean_me}";
+    else
+      msg "'${clean_me}' does not appear to be a directory!";
+      exit 1;
+    fi;
 }
 
 print_help() {
+    echo "################################################################################"
     echo "Usage: ./examples/yt-wsp.sh <video_url>"
-    echo "See configurable env variables in the script"
-    echo "This will produce an MP4 muxed file called res.mp4 in the working directory"
-    echo "Requirements: ffmpeg yt-dlp whisper"
-    echo "Whisper needs to be built into the main binary with make, then you can rename it to something like 'whisper' and add it to your PATH for convenience."
-    echo "E.g. in the root of Whisper.cpp, run: 'make && cp ./main /usr/local/bin/whisper'"
+    echo "# See configurable env variables in the script; there are many!"
+    echo "# This script will produce an MP4 muxed file in the working directory; it will"
+    echo "# be named for the title and id of the video."
+    echo "# passing in https://youtu.be/VYJtb2YXae8 produces a file named";
+    echo "# 'Why_we_all_need_subtitles_now-VYJtb2YXae8-res.mp4'"
+    echo "# Requirements: ffmpeg yt-dlp whisper.cpp"
+    echo "################################################################################"
 }
 
 check_requirements() {
     if ! command -v ffmpeg &>/dev/null; then
-        echo "ffmpeg is required (https://ffmpeg.org)."
+        echo "ffmpeg is required: https://ffmpeg.org";
         exit 1
-    fi
+    fi;
 
     if ! command -v yt-dlp &>/dev/null; then
-        echo "yt-dlp is required (https://github.com/yt-dlp/yt-dlp)."
-        exit 1
-    fi
+        echo "yt-dlp is required: https://github.com/yt-dlp/yt-dlp";
+        exit 1;
+    fi;
+
+    if ! command -v "${WHISPER_EXECUTABLE}" &>/dev/null; then
+        echo "The C++ implementation of Whisper is required: https://github.com/ggerganov/whisper.cpp"
+        echo "Sample usage:";
+        echo "";
+        echo "  git clone https://github.com/ggerganov/whisper.cpp";
+        echo "  cd whisper.cpp";
+        echo "  make";
+        echo "  ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890";
+        echo "";
+        exit 1;
+    fi;
 
-    if ! command -v "$WHISPER_EXECUTABLE" &>/dev/null; then
-        WHISPER_EXECUTABLE="./main"
-        if ! command -v "$WHISPER_EXECUTABLE" &>/dev/null; then
-            echo "Whisper is required (https://github.com/ggerganov/whisper.cpp):"
-            echo "Sample usage:"
-            echo ""
-            echo "  git clone https://github.com/ggerganov/whisper.cpp"
-            echo "  cd whisper.cpp"
-            echo "  make"
-            echo "  ./examples/yt-wsp.sh https://www.youtube.com/watch?v=1234567890"
-            echo ""
-            exit 1
-        fi
-    fi
 }
 
-if [[ $# -lt 1 ]]; then
-    print_help
-    exit 1
+if [[ "${#}" -lt 1 ]]; then
+    print_help;
+    exit 1;
 fi
 
-if [[ "$1" == "help" ]]; then
-    print_help
-    exit 0
+if [[ "${1##-*}" == "help" ]]; then
+    print_help;
+    exit 0;
 fi
 
-temp_dir="tmp"
-source_url="$1"
+check_requirements;
 
-check_requirements
+################################################################################
+# create a temporary directory to work in
+# set the temp_dir and temp_filename variables
+################################################################################
+temp_dir="$(mktemp -d ${SCRIPT_DIR}/tmp.XXXXXX)";
+temp_filename="${temp_dir}/yt-dlp-filename";
 
-msg "Downloading VOD..."
+################################################################################
+# for now we only take one argument
+# TODO: a for loop
+################################################################################
+source_url="${1}"
+title_name="";
 
-# Optionally add --cookies-from-browser BROWSER[+KEYRING][:PROFILE][::CONTAINER] for members only VODs
+msg "Downloading VOD...";
+
+################################################################################
+# Download the video, put the dynamic output filename into a variable.
+# Optionally add --cookies-from-browser BROWSER[+KEYRING][:PROFILE][::CONTAINER]
+# for videos only available to logged-in users.
+################################################################################
 yt-dlp \
     -f "bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best" \
+    -o "${temp_dir}/%(title)s-%(id)s.vod.mp4" \
+    --print-to-file "%(filename)s" "${temp_filename}" \
+    --no-simulate \
+    --no-write-auto-subs \
+    --restrict-filenames \
     --embed-thumbnail \
     --embed-chapters \
     --xattrs \
-    "${source_url}" -o "${temp_dir}/vod.mp4"
+    "${source_url}";
 
-msg "Extracting audio and resampling..."
+title_name="$(xargs basename -s .vod.mp4 < ${temp_filename})";
 
-ffmpeg -i "${temp_dir}/vod.mp4" \
+msg "Extracting audio and resampling...";
+
+ffmpeg -i "${temp_dir}/${title_name}.vod.mp4"  \
     -hide_banner \
+    -vn \
     -loglevel error \
     -ar 16000 \
     -ac 1 \
-    -c:a \
-    pcm_s16le -y "vod-resampled.wav"
+    -c:a pcm_s16le \
+    -y \
+    "${temp_dir}/${title_name}.vod-resampled.wav";
 
-msg "Transcribing to subtitle file..."
-msg "Whisper specified at: ${WHISPER_EXECUTABLE}"
+msg "Transcribing to subtitle file...";
+msg "Whisper specified at: '${WHISPER_EXECUTABLE}'";
 
-$WHISPER_EXECUTABLE \
+"${WHISPER_EXECUTABLE}" \
     -m "${MODEL_PATH}" \
     -l "${WHISPER_LANG}" \
-    -f "vod-resampled.wav" \
-    -t 8 \
+    -f "${temp_dir}/${title_name}.vod-resampled.wav" \
+    -t "${WHISPER_THREAD_COUNT}" \
     -osrt \
-    --translate
+    --translate;
 
-msg "Embedding subtitle track..."
+msg "Embedding subtitle track...";
 
-ffmpeg -i "${temp_dir}/vod.mp4" \
+ffmpeg -i "${temp_dir}/${title_name}.vod.mp4" \
     -hide_banner \
     -loglevel error \
-    -i "vod-resampled.wav.srt" \
+    -i "${temp_dir}/${title_name}.vod-resampled.wav.srt" \
     -c copy \
     -c:s mov_text \
-    -y res.mp4
+    -y "${title_name}-res.mp4";
 
-cleanup
+#cleanup "${temp_dir}";
 
-msg "Done! Your finished file is ready: res.mp4"
+msg "Done! Your finished file is ready: ${title_name}-res.mp4";

extra/bench-all.sh

@@ -12,6 +12,18 @@ fi
 
 models=( "tiny" "base" "small" "medium" "large" )
 
+printf "\n"
+printf "Running memcpy benchmark with 1 thread\n"
+printf "\n"
+
+./bench -w 1 -t 1 2>&1
+
+printf "\n"
+printf "Running ggml_mul_mat benchmark with $n_threads threads\n"
+printf "\n"
+
+./bench -w 2 -t $n_threads 2>&1
+
 printf "\n"
 printf "Running benchmark for all models\n"
 printf "This can take a while!\n"
@@ -56,4 +68,3 @@ for model in "${models[@]}"; do
 
     printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
 done
-

ggml.c

@@ -84,7 +84,7 @@ typedef void* thread_ret_t;
 #define GGML_GELU_FP16
 
 #define GGML_SOFT_MAX_UNROLL 4
-#define GGML_VEC_DOT_UNROLL  4
+#define GGML_VEC_DOT_UNROLL  2
 
 #ifdef GGML_USE_ACCELERATE
 // uncomment to use vDSP for soft max computation
@@ -339,8 +339,12 @@ int64_t ggml_cycles_per_ms(void) {
 #if defined(__cpp_lib_hardware_interference_size)
 #define CACHE_LINE_SIZE hardware_destructive_interference_size
 #else
+#if defined(__POWER9_VECTOR__)
+#define CACHE_LINE_SIZE 128
+#else
 #define CACHE_LINE_SIZE 64
 #endif
+#endif
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
@@ -609,9 +613,12 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 #define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ?                   \
   vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \
   vec_extract_fp32_from_shortl(vec_xl(0, p))
-#define GGML_F16_VEC_STORE(p, r, i)                                      \
-  if (i & 0x1)                                                           \
-    vec_xst(vec_pack_to_short_fp32(r[i], r[i - 1]), 0, p - GGML_F16_EPR)
+#define GGML_ENDIAN_BYTE(i) ((unsigned char *)&(uint16_t){1})[i]
+#define GGML_F16_VEC_STORE(p, r, i)                             \
+  if (i & 0x1)                                                  \
+    vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)],  \
+                                   r[i - GGML_ENDIAN_BYTE(0)]), \
+            0, p - GGML_F16_EPR)
 
 #elif defined(__wasm_simd128__)
 
@@ -923,9 +930,9 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t
 inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
     ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
 
-    const ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL] = { xv };
+    ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL];
 
-    for (int i = 1; i < GGML_VEC_DOT_UNROLL; ++i) {
+    for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
         x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
     }
 
@@ -1109,8 +1116,8 @@ inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
     ggml_float sum = 0.0;
     for (int i = 0; i < n; ++i) {
         sum += x[i];
-        *s += sum;
     }
+    *s = sum;
 #else
     vDSP_sve(x, 1, s, n);
 #endif
@@ -1251,7 +1258,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
 //
 
 struct ggml_object {
-    size_t offset;
+    size_t offs;
     size_t size;
 
     struct ggml_object * next;
@@ -1277,6 +1284,9 @@ struct ggml_context {
 
     struct ggml_object * objects_begin;
     struct ggml_object * objects_end;
+
+    struct ggml_scratch scratch;
+    struct ggml_scratch scratch_save;
 };
 
 struct ggml_context_container {
@@ -1339,7 +1349,7 @@ inline static void ggml_critical_section_end(void) {
 
 void ggml_print_object(const struct ggml_object * obj) {
     GGML_PRINT(" - ggml_object: offset = %zu, size = %zu, next = %p\n",
-            obj->offset, obj->size, (const void *) obj->next);
+            obj->offs, obj->size, (const void *) obj->next);
 }
 
 void ggml_print_objects(const struct ggml_context * ctx) {
@@ -1535,12 +1545,14 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
     }
 
     *ctx = (struct ggml_context) {
-        .mem_size         = params.mem_size,
-        .mem_buffer       = params.mem_buffer ? params.mem_buffer : malloc(params.mem_size),
-        .mem_buffer_owned = params.mem_buffer ? false : true,
-        .n_objects        = 0,
-        .objects_begin    = NULL,
-        .objects_end      = NULL,
+        /*.mem_size         =*/ params.mem_size,
+        /*.mem_buffer       =*/ params.mem_buffer ? params.mem_buffer : malloc(params.mem_size),
+        /*.mem_buffer_owned =*/ params.mem_buffer ? false : true,
+        /*.n_objects        =*/ 0,
+        /*.objects_begin    =*/ NULL,
+        /*.objects_end      =*/ NULL,
+        /*.scratch          =*/ { 0, 0, NULL, },
+        /*.scratch_save     =*/ { 0, 0, NULL, },
     };
 
     ggml_assert_aligned(ctx->mem_buffer);
@@ -1563,7 +1575,7 @@ void ggml_free(struct ggml_context * ctx) {
             g_state.contexts[i].used = false;
 
             GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n",
-                    __func__, i, ctx->n_objects, ctx->objects_end->offset + ctx->objects_end->size);
+                    __func__, i, ctx->n_objects, ctx->objects_end->offs + ctx->objects_end->size);
 
             if (ctx->mem_buffer_owned) {
                 free(ctx->mem_buffer);
@@ -1582,7 +1594,15 @@ void ggml_free(struct ggml_context * ctx) {
 }
 
 size_t ggml_used_mem(const struct ggml_context * ctx) {
-    return ctx->objects_end->offset + ctx->objects_end->size;
+    return ctx->objects_end->offs + ctx->objects_end->size;
+}
+
+size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) {
+    const size_t result = ctx->scratch.data ? ctx->scratch.offs : 0;
+
+    ctx->scratch = scratch;
+
+    return result;
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -1596,9 +1616,9 @@ struct ggml_tensor * ggml_new_tensor_impl(
     // always insert objects at the end of the context's memory pool
     struct ggml_object * obj_cur = ctx->objects_end;
 
-    const size_t cur_offset = obj_cur == NULL ? 0 : obj_cur->offset;
-    const size_t cur_size   = obj_cur == NULL ? 0 : obj_cur->size;
-    const size_t cur_end    = cur_offset + cur_size;
+    const size_t cur_offs = obj_cur == NULL ? 0 : obj_cur->offs;
+    const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size;
+    const size_t cur_end  = cur_offs + cur_size;
 
     size_t size_needed = 0;
 
@@ -1609,25 +1629,52 @@ struct ggml_tensor * ggml_new_tensor_impl(
         }
         // align to GGML_MEM_ALIGN
         size_needed = ((size_needed + GGML_MEM_ALIGN - 1)/GGML_MEM_ALIGN)*GGML_MEM_ALIGN;
-
-    }
-    size_needed += sizeof(struct ggml_tensor);
-
-    if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
-        GGML_PRINT("%s: not enough space in the context's memory pool\n", __func__);
-        assert(false);
-        return NULL;
     }
 
     char * const mem_buffer = ctx->mem_buffer;
-
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
-    *obj_new = (struct ggml_object) {
-        .offset = cur_end + GGML_OBJECT_SIZE,
-        .size   = size_needed,
-        .next   = NULL,
-    };
+    if (ctx->scratch.data == NULL || data != NULL) {
+        size_needed += sizeof(struct ggml_tensor);
+
+        if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
+            GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
+                    __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
+            assert(false);
+            return NULL;
+        }
+
+        *obj_new = (struct ggml_object) {
+            .offs = cur_end + GGML_OBJECT_SIZE,
+            .size = size_needed,
+            .next = NULL,
+        };
+    } else {
+        if (ctx->scratch.offs + size_needed > ctx->scratch.size) {
+            GGML_PRINT("%s: not enough space in the scratch memory\n", __func__);
+            assert(false);
+            return NULL;
+        }
+
+        if (cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE > ctx->mem_size) {
+            GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
+                    __func__, cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE, ctx->mem_size);
+            assert(false);
+            return NULL;
+        }
+
+        data = (char * const) ctx->scratch.data + ctx->scratch.offs;
+
+        *obj_new = (struct ggml_object) {
+            .offs = cur_end + GGML_OBJECT_SIZE,
+            .size = sizeof(struct ggml_tensor),
+            .next = NULL,
+        };
+
+        //printf("scratch offs = %zu, size_needed = %zu\n", ctx->scratch.offs, size_needed);
+
+        ctx->scratch.offs += size_needed;
+    }
 
     if (obj_cur != NULL) {
         obj_cur->next = obj_new;
@@ -1638,9 +1685,9 @@ struct ggml_tensor * ggml_new_tensor_impl(
 
     ctx->objects_end = obj_new;
 
-    //GGML_PRINT_DEBUG("%s: inserted new object at %zu\n", __func__, cur_end);
+    //printf("%s: inserted new object at %zu, size = %zu\n", __func__, cur_end, obj_new->size);
 
-    struct ggml_tensor * const result = (struct ggml_tensor *)(mem_buffer + obj_new->offset);
+    struct ggml_tensor * const result = (struct ggml_tensor *)(mem_buffer + obj_new->offs);
 
     ggml_assert_aligned(result);
 
@@ -1683,7 +1730,7 @@ struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int* ne) {
+        const int * ne) {
     return ggml_new_tensor_impl(ctx, type, n_dims, ne, NULL);
 }
 
@@ -1725,16 +1772,26 @@ struct ggml_tensor * ggml_new_tensor_4d(
 }
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
+    ctx->scratch_save = ctx->scratch;
+    ctx->scratch.data = NULL;
+
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
 
+    ctx->scratch = ctx->scratch_save;
+
     ggml_set_i32(result, value);
 
     return result;
 }
 
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
+    ctx->scratch_save = ctx->scratch;
+    ctx->scratch.data = NULL;
+
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
 
+    ctx->scratch = ctx->scratch_save;
+
     ggml_set_f32(result, value);
 
     return result;
@@ -2343,7 +2400,7 @@ struct ggml_tensor * ggml_repeat(
     result->op   = GGML_OP_REPEAT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src0 = a;
-    result->src1 = NULL;
+    result->src1 = b;
 
     return result;
 }
@@ -2959,9 +3016,7 @@ struct ggml_tensor * ggml_diag_mask_inf(
     // TODO: when implement backward, fix this:
     //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
-    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
-    ((int32_t *) b->data)[0] = n_past;
+    struct ggml_tensor * b = ggml_new_i32(ctx, n_past);
 
     result->op   = GGML_OP_DIAG_MASK_INF;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -3724,8 +3779,6 @@ static void ggml_compute_forward_sum_f32(
     assert(ggml_is_scalar(dst));
     assert(src0->nb[0] == sizeof(float));
 
-    *(float *) (dst->data) = 0.0f;
-
     const int ne00 = src0->ne[0];
     const int ne01 = src0->ne[1];
     const int ne02 = src0->ne[2];
@@ -3811,8 +3864,6 @@ static void ggml_compute_forward_mean_f32(
     for (int i03 = 0; i03 < ne03; i03++) {
         for (int i02 = 0; i02 < ne02; i02++) {
             for (int i01 = 0; i01 < ne01; i01++) {
-                *(float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3) = 0.0f;
-
                 ggml_vec_sum_f32(ne00,
                         (float *) ((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3),
                         (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03));
@@ -4297,7 +4348,9 @@ static bool ggml_compute_forward_mul_mat_use_blas(
     const int ne1 = dst->ne[1];
 
     // TODO: find the optimal values for these
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ne0 >= 32 && ne1 >= 32 && ne10 >= 32) {
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && (
+             (ne0 >= 32 && ne1  >= 32   && ne10 >= 32)
+            )) {
         //printf("BLAS: %d %d %d\n", ne0, ne1, ne10);
         return true;
     }
@@ -4377,7 +4430,9 @@ static void ggml_compute_forward_mul_mat_f32(
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         GGML_ASSERT(nb10 == sizeof(float));
 
-        if (params->ith != 0) return;
+        if (params->ith != 0) {
+            return;
+        }
 
         if (params->type == GGML_TASK_INIT) {
             return;
@@ -4620,7 +4675,9 @@ static void ggml_compute_forward_mul_mat_f16_f32(
     if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
         GGML_ASSERT(nb10 == sizeof(float));
 
-        if (params->ith != 0) return;
+        if (params->ith != 0) {
+            return;
+        }
 
         if (params->type == GGML_TASK_INIT) {
             return;
@@ -4791,7 +4848,7 @@ static void ggml_compute_forward_mul_mat_f16_f32(
             }
         }
     } else {
-        // parallelize by src1 columns using ggml_vec_mad_f32
+        // parallelize by src1 columns using ggml_vec_mad_f16
         // each thread has its own work data
         // during FINALIZE we accumulate all work data into dst
 
@@ -6158,40 +6215,37 @@ static void ggml_compute_forward_flash_attn_f16(
             S[i] = -INFINITY;
         }
 
-        // looks like unrolling here does not help
-#if 1
-        for (int ic = 0; ic < nek1; ++ic) {
-            // k indices
-            const int ik3 = iq3;
-            const int ik2 = iq2;
-            const int ik1 = ic;
+        if (GGML_VEC_DOT_UNROLL > 2 || nek1 % GGML_VEC_DOT_UNROLL != 0) {
+            for (int ic = 0; ic < nek1; ++ic) {
+                // k indices
+                const int ik3 = iq3;
+                const int ik2 = iq2;
+                const int ik1 = ic;
 
-            // S indices
-            const int i1 = ik1;
+                // S indices
+                const int i1 = ik1;
 
-            ggml_vec_dot_f16(neq0,
-                    S + i1,
-                    (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                    (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                ggml_vec_dot_f16(neq0,
+                        S + i1,
+                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+            }
+        } else {
+            for (int ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
+                // k indices
+                const int ik3 = iq3;
+                const int ik2 = iq2;
+                const int ik1 = ic;
+
+                // S indices
+                const int i1 = ik1;
+
+                ggml_vec_dot_f16_unroll(neq0, nbk1,
+                        S + i1,
+                        ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+            }
         }
-#else
-        GGML_ASSERT(nek1 % GGML_VEC_DOT_UNROLL == 0);
-
-        for (int ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
-            // k indices
-            const int ik3 = iq3;
-            const int ik2 = iq2;
-            const int ik1 = ic;
-
-            // S indices
-            const int i1 = ik1;
-
-            ggml_vec_dot_f16_unroll(neq0, nbk1,
-                    S + i1,
-                                    ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                    (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-        }
-#endif
 
         // scale
         ggml_vec_scale_f32(nek1, S, scale);
@@ -6261,18 +6315,30 @@ static void ggml_compute_forward_flash_attn_f16(
             S16[i] = GGML_FP32_TO_FP16(S[i]);
         }
 
-        GGML_ASSERT(nev1 % GGML_VEC_DOT_UNROLL == 0);
+        if (GGML_VEC_DOT_UNROLL == 1 || (nev1 % GGML_VEC_DOT_UNROLL != 0)) {
+            for (int ic = 0; ic < nev1; ++ic) {
+                // dst indices
+                const int i1 = iq1;
+                const int i2 = iq2;
+                const int i3 = iq3;
 
-        for (int ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
-            // dst indices
-            const int i1 = iq1;
-            const int i2 = iq2;
-            const int i3 = iq3;
+                ggml_vec_dot_f16(nek1,
+                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
+                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
+                        S16);
+            }
+        } else {
+            for (int ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
+                // dst indices
+                const int i1 = iq1;
+                const int i2 = iq2;
+                const int i3 = iq3;
 
-            ggml_vec_dot_f16_unroll(nek1, nbv1,
-                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
-                              ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
-                    S16);
+                ggml_vec_dot_f16_unroll(nek1, nbv1,
+                        (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2  + i3*nb3)),
+                        ((char *) v->data   + (         ic*nbv1 + i2*nbv2 + i3*nbv3)),
+                        S16);
+            }
         }
     }
 }
@@ -7049,7 +7115,7 @@ struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cg
 #ifdef __APPLE__
 
 //#include <os/lock.h>
-
+//
 //typedef os_unfair_lock ggml_lock_t;
 //
 //#define ggml_lock_init(x)    UNUSED(x)
@@ -7156,6 +7222,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             if (state->params.ith < state->params.nth) {
                 ggml_compute_forward(&state->params, state->node);
             }
+
             state->node = NULL;
         } else {
             break;
@@ -7200,6 +7267,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                 .node   = NULL,
                 .shared = &state_shared,
             };
+
             int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
             assert(rc == 0);
             UNUSED(rc);
@@ -7268,8 +7336,12 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                                 node->src1->type == GGML_TYPE_F32) {
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
                                 if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
-                                    node->n_tasks = 1;
+                                    node->n_tasks = 1; // TODO: this actually is doing nothing
+                                                       //       the threads are still spinning
                                     cur = sizeof(float)*(node->src0->ne[0]*node->src0->ne[1]);
+                                    //printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
+                                    //printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
+                                    //printf("cur = %zu\n", cur);
                                 } else {
                                     cur = sizeof(ggml_fp16_t)*ggml_nelements(node->src1);
                                 }
@@ -8445,6 +8517,195 @@ enum ggml_opt_result ggml_opt(
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void ggml_svd_reduce_dims(
+        int ne0,
+        int ne1,
+        float * a,
+        int nd) {
+    int n = ne1;
+    int m = ne0;
+
+    float * A = a;
+    float * A0 = (float *) malloc(n * m * sizeof(float));
+
+    // average vector
+    //float * M = (float *) malloc(m * sizeof(float));
+
+    //{
+    //    for (int j = 0; j < m; ++j) {
+    //        M[j] = 0.0f;
+    //    }
+    //    for (int i = 0; i < n; ++i) {
+    //        for (int j = 0; j < m; ++j) {
+    //            M[j] += A[i * m + j];
+    //        }
+    //    }
+    //    for (int j = 0; j < m; ++j) {
+    //        M[j] /= (float) n;
+    //    }
+    //}
+
+    //// subtract average vector
+    //for (int i = 0; i < n; ++i) {
+    //    for (int j = 0; j < m; ++j) {
+    //        A[i * m + j] -= M[j];
+    //    }
+    //}
+
+    //free(M);
+
+    memcpy(A0, A, n * m * sizeof(float));
+
+    // print A
+    //printf("A:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < m; ++j) {
+    //        printf("%9.5f ", A[i * m + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    // SVD
+    // A = U * S * V^T
+
+    float * U = (float *) malloc(n * m * sizeof(float));
+    float * S = (float *) malloc(n * sizeof(float));
+    float * V = (float *) malloc(n * n * sizeof(float));
+
+    int lda = m;
+    int ldu = m;
+    int ldvt = n;
+
+    float work_size;
+    int lwork = -1;
+    int info = 0;
+
+    sgesvd_("S", "S", &m, &n, A, &lda, S, U, &ldu, V, &ldvt, &work_size, &lwork, &info);
+
+    lwork = (int) work_size;
+
+    //printf("work_size = %f, info = %d, lwork = %d\n", work_size, info, lwork);
+
+    float * work = (float *) malloc(lwork * sizeof(float));
+
+    sgesvd_("S", "S", &m, &n, A, &lda, S, U, &ldu, V, &ldvt, work, &lwork, &info);
+
+    free(work);
+
+    // print U
+    //printf("U:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < m; ++j) {
+    //        printf("%9.5f ", U[i * m + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    // normalize S
+    {
+        double sum = 0.0;
+        for (int i = 0; i < n; ++i) {
+            sum += S[i];
+        }
+        sum *= sqrt((double) m);
+        for (int i = 0; i < n; ++i) {
+            S[i] /= sum;
+        }
+    }
+
+    // print S
+    printf("S:\n");
+    for (int i = 0; i < n; ++i) {
+        printf("- %d = %9.5f\n", i, S[i]);
+    }
+    printf("\n");
+
+    // print V
+    //printf("V:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < n; ++j) {
+    //        printf("%9.5f ", V[i * n + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    // print A
+    //printf("A:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < m; ++j) {
+    //        printf("%9.5f ", A[i * m + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    // compute singular vectors in U
+    for (int i = 0; i < n; ++i) {
+        for (int j = 0; j < m; ++j) {
+            U[i * m + j] *= S[i];
+        }
+    }
+
+    // normalize U
+    for (int i = 0; i < n; ++i) {
+        double sum = 0.0;
+        for (int j = 0; j < m; ++j) {
+            sum += U[i * m + j] * U[i * m + j];
+        }
+        sum = sqrt(sum);
+        for (int j = 0; j < m; ++j) {
+            U[i * m + j] /= sum*sqrt((double) m);
+        }
+    }
+
+    // print U
+    //printf("U:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < m; ++j) {
+    //        printf("%9.5f ", U[i * m + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    // project A0 onto U
+    for (int i = 0; i < n; ++i) {
+        for (int j = 0; j < nd; ++j) {
+            A[i * nd + j] = 0.0f;
+            //if (j == 0) continue;
+            for (int k = 0; k < m; ++k) {
+                A[i * nd + j] += A0[i * m + k] * U[j * m + k];
+            }
+        }
+    }
+
+    // print A
+    //printf("A:\n");
+    //for (int i = 0; i < n; ++i) {
+    //    printf("col %d : ", i);
+    //    for (int j = 0; j < n; ++j) {
+    //        printf("%9.5f ", A[i * n + j]);
+    //    }
+    //    printf("\n");
+    //}
+    //printf("\n");
+
+    free(U);
+    free(S);
+    free(V);
+    free(A0);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
 int ggml_cpu_has_avx(void) {
 #if defined(__AVX__)
     return 1;

ggml.h

@@ -301,6 +301,13 @@ struct ggml_cgraph {
     int64_t perf_time_us;
 };
 
+// scratch buffer
+struct ggml_scratch {
+    size_t offs;
+    size_t size;
+    void * data;
+};
+
 struct ggml_init_params {
     // memory pool
     size_t mem_size;   // bytes
@@ -327,6 +334,8 @@ void ggml_free(struct ggml_context * ctx);
 
 size_t ggml_used_mem(const struct ggml_context * ctx);
 
+size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch);
+
 struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
@@ -717,6 +726,16 @@ enum ggml_opt_result ggml_opt(
         struct ggml_opt_params params,
         struct ggml_tensor * f);
 
+//
+// Temp stuff
+//
+
+void ggml_svd_reduce_dims(
+        int ne0,
+        int ne1,
+        float * a,
+        int nd);
+
 //
 // system info
 //

whisper.h

@@ -1,6 +1,7 @@
 #ifndef WHISPER_H
 #define WHISPER_H
 
+#include <stddef.h>
 #include <stdint.h>
 #include <stdbool.h>
 
@@ -40,7 +41,7 @@ extern "C" {
     //
     //     ...
     //
-    //     struct whisper_context * ctx = whisper_init("/path/to/ggml-base.en.bin");
+    //     struct whisper_context * ctx = whisper_init_from_file("/path/to/ggml-base.en.bin");
     //
     //     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
     //         fprintf(stderr, "failed to process audio\n");
@@ -73,6 +74,7 @@ extern "C" {
         whisper_token tid; // forced timestamp token id
 
         float p;           // probability of the token
+        float plog;        // log probability of the token
         float pt;          // probability of the timestamp token
         float ptsum;       // sum of probabilities of all timestamp tokens
 
@@ -84,9 +86,20 @@ extern "C" {
         float vlen;        // voice length of the token
     } whisper_token_data;
 
-    // Allocates all memory needed for the model and loads the model from the given file.
-    // Returns NULL on failure.
-    WHISPER_API struct whisper_context * whisper_init(const char * path_model);
+    typedef struct whisper_model_loader {
+        void * context;
+
+        size_t (*read)(void * ctx, void * output, size_t read_size);
+        bool    (*eof)(void * ctx);
+        void  (*close)(void * ctx);
+    } whisper_model_loader;
+
+    // Various functions for loading a ggml whisper model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    WHISPER_API struct whisper_context * whisper_init_from_file(const char * path_model);
+    WHISPER_API struct whisper_context * whisper_init_from_buffer(void * buffer, size_t buffer_size);
+    WHISPER_API struct whisper_context * whisper_init(struct whisper_model_loader * loader);
 
     // Frees all memory allocated by the model.
     WHISPER_API void whisper_free(struct whisper_context * ctx);
@@ -100,6 +113,16 @@ extern "C" {
                                int   n_samples,
                                int   n_threads);
 
+    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2. 
+    // The resulting spectrogram is stored inside the provided whisper context.
+    // Returns 0 on success
+    WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
+        struct whisper_context* ctx,
+        const float* samples,
+        int   n_samples,
+        int   n_threads);
+
+
     // 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
@@ -124,6 +147,7 @@ extern "C" {
     // tokens + n_tokens is the provided context for the decoder.
     // n_past is the number of tokens to use from previous decoder calls.
     // Returns 0 on success
+    // TODO: add support for multiple decoders
     WHISPER_API int whisper_decode(
             struct whisper_context * ctx,
                const whisper_token * tokens,
@@ -131,14 +155,6 @@ extern "C" {
                                int   n_past,
                                int   n_threads);
 
-    // Token sampling methods.
-    // These are provided for convenience and can be used after each call to whisper_decode().
-    // You can also implement your own sampling method using the whisper_get_probs() function.
-    // whisper_sample_best() returns the token with the highest probability
-    // whisper_sample_timestamp() returns the most probable timestamp token
-    WHISPER_API whisper_token_data whisper_sample_best(struct whisper_context * ctx);
-    WHISPER_API whisper_token_data whisper_sample_timestamp(struct whisper_context * ctx, bool is_initial);
-
     // 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
@@ -180,8 +196,11 @@ extern "C" {
     WHISPER_API int whisper_n_audio_ctx    (struct whisper_context * ctx);
     WHISPER_API int whisper_is_multilingual(struct whisper_context * ctx);
 
-    // The probabilities for the next token
-    WHISPER_API float * whisper_get_probs(struct whisper_context * ctx);
+    // Token logits obtained from the last call to whisper_decode()
+    // The logits for the last token are stored in the last row
+    // Rows: n_tokens
+    // Cols: n_vocab
+    WHISPER_API float * whisper_get_logits(struct whisper_context * ctx);
 
     // Token Id -> String. Uses the vocabulary in the provided context
     WHISPER_API const char * whisper_token_to_str(struct whisper_context * ctx, whisper_token token);
@@ -210,8 +229,8 @@ extern "C" {
 
     // Available sampling strategies
     enum whisper_sampling_strategy {
-        WHISPER_SAMPLING_GREEDY,      // Always select the most probable token
-        WHISPER_SAMPLING_BEAM_SEARCH, // TODO: not implemented yet!
+        WHISPER_SAMPLING_GREEDY,      // similar to OpenAI's GreefyDecoder
+        WHISPER_SAMPLING_BEAM_SEARCH, // similar to OpenAI's BeamSearchDecoder
     };
 
     // Text segment callback
@@ -231,30 +250,32 @@ extern "C" {
         enum whisper_sampling_strategy strategy;
 
         int n_threads;
-        int n_max_text_ctx;
+        int n_max_text_ctx;     // max tokens to use from past text as prompt for the decoder
         int offset_ms;          // start offset in ms
         int duration_ms;        // audio duration to process in ms
 
         bool translate;
-        bool no_context;
+        bool no_context;        // do not use past transcription (if any) as initial prompt for the decoder
         bool single_segment;    // force single segment output (useful for streaming)
-        bool print_special;
-        bool print_progress;
-        bool print_realtime;
-        bool print_timestamps;
+        bool print_special;     // print special tokens (e.g. <SOT>, <EOT>, <BEG>, etc.)
+        bool print_progress;    // print progress information
+        bool print_realtime;    // print results from within whisper.cpp (avoid it, use callback instead)
+        bool print_timestamps;  // print timestamps for each text segment when printing realtime
 
         // [EXPERIMENTAL] token-level timestamps
         bool  token_timestamps; // enable token-level timestamps
         float thold_pt;         // timestamp token probability threshold (~0.01)
         float thold_ptsum;      // timestamp token sum probability threshold (~0.01)
         int   max_len;          // max segment length in characters
+        bool  split_on_word;    // split on word rather than on token (when used with max_len)
         int   max_tokens;       // max tokens per segment (0 = no limit)
 
         // [EXPERIMENTAL] speed-up techniques
+        // note: these can significantly reduce the quality of the output
         bool speed_up;          // speed-up the audio by 2x using Phase Vocoder
         int  audio_ctx;         // overwrite the audio context size (0 = use default)
 
-        // tokens to provide the whisper model as initial prompt
+        // tokens to provide to the whisper decoder as initial prompt
         // these are prepended to any existing text context from a previous call
         const whisper_token * prompt_tokens;
         int prompt_n_tokens;
@@ -262,19 +283,36 @@ extern "C" {
         // for auto-detection, set to nullptr, "" or "auto"
         const char * language;
 
+        // common decoding parameters:
+        bool suppress_blank;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L89
+        bool suppress_non_speech_tokens; // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
+
+        float temperature;      // initial decoding temperature, ref: https://ai.stackexchange.com/a/32478
+        float max_initial_ts;   // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/decoding.py#L97
+        float length_penalty;   // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L267
+
+        // fallback parameters
+        // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L274-L278
+        float temperature_inc;
+        float entropy_thold;    // similar to OpenAI's "compression_ratio_threshold"
+        float logprob_thold;
+        float no_speech_thold;  // TODO: not implemented
+
         struct {
-            int n_past;
+            int best_of;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264
         } greedy;
 
         struct {
-            int n_past;
-            int beam_width;
-            int n_best;
+            int beam_size;  // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L265
+
+            float patience; // TODO: not implemented, ref: https://arxiv.org/pdf/2204.05424.pdf
         } beam_search;
 
+        // called for every newly generated text segment
         whisper_new_segment_callback new_segment_callback;
         void * new_segment_callback_user_data;
 
+        // called each time before the encoder starts
         whisper_encoder_begin_callback encoder_begin_callback;
         void * encoder_begin_callback_user_data;
     };
@@ -303,6 +341,9 @@ extern "C" {
     // A segment can be a few words, a sentence, or even a paragraph.
     WHISPER_API int whisper_full_n_segments(struct whisper_context * ctx);
 
+    // Language id associated with the current context
+    WHISPER_API int whisper_full_lang_id(struct whisper_context * ctx);
+
     // Get the start and end time of the specified segment.
     WHISPER_API int64_t whisper_full_get_segment_t0(struct whisper_context * ctx, int i_segment);
     WHISPER_API int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment);
@@ -324,6 +365,17 @@ extern "C" {
     // Get the probability of the specified token in the specified segment.
     WHISPER_API float whisper_full_get_token_p(struct whisper_context * ctx, int i_segment, int i_token);
 
+    ////////////////////////////////////////////////////////////////////////////
+
+    // Temporary helpers needed for exposing ggml interface
+
+    WHISPER_API int whisper_bench_memcpy(int n_threads);
+    WHISPER_API int whisper_bench_ggml_mul_mat(int n_threads);
+
+    // Temporary experimental API
+
+    WHISPER_API void whisper_full_cluster_segments(struct whisper_context * ctx);
+
 #ifdef __cplusplus
 }
 #endif