whisper : account speed_up flag for short audio (close #405 )

Update README.md
bench.wasm : print system info
2025-06-25 09:31:44 +00:00 · 2023-01-15 12:42:15 +02:00 · 2023-01-15 11:36:20 +02:00 · 2023-01-15 11:34:03 +02:00 · 2023-01-15 11:29:57 +02:00 · 2023-01-14 19:20:47 +02:00
63 changed files with 3558 additions and 1912 deletions
--- a/.github/workflows/bindings.yml
+++ b/.github/workflows/bindings.yml
@ -0,0 +1,17 @@
+name: Bindings Tests
+on:
+  push:
+    paths:
+      - bindings/go/**
+
+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
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -235,3 +235,33 @@ jobs:
              with:
                name: whisper-blas-bin-${{ matrix.arch }}
                path: build/bin/${{ matrix.build }}
+
+    emscripten:
+        runs-on: ubuntu-latest
+
+        strategy:
+            matrix:
+                build: [Release]
+
+        steps:
+            - name: Clone
+              uses: actions/checkout@v1
+
+            - name: Dependencies
+              run: |
+                wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz
+                tar -xvf master.tar.gz
+                emsdk-master/emsdk update
+                emsdk-master/emsdk install latest
+                emsdk-master/emsdk activate latest
+
+            - name: Configure
+              run: echo "tmp"
+
+            - name: Build
+              run: |
+                pushd emsdk-master
+                source ./emsdk_env.sh
+                popd
+                emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+                make
--- a/.gitignore
+++ b/.gitignore
@ -8,6 +8,7 @@ build/
 build-em/
 build-debug/
 build-release/
+build-static/
 build-sanitize-addr/
 build-sanitize-thread/

@ -18,6 +19,7 @@ build-sanitize-thread/
 /bench

 sync.sh
+libwhisper.a
 libwhisper.so
 compile_commands.json

--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -2,14 +2,15 @@ cmake_minimum_required (VERSION 3.0)

 project(whisper.cpp VERSION 1.0.4)

-set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
+# Add path to modules
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
-set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")

 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
    set(WHISPER_STANDALONE ON)
-    include(cmake/GitVars.cmake)
-    include(cmake/BuildTypes.cmake)
+    include(GitVars)
+    include(BuildTypes)

    # configure project version
    if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
@ -52,6 +53,7 @@ if (APPLE)
    option(WHISPER_NO_ACCELERATE       "whisper: disable Accelerate framework" OFF)
    option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
    option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
+    option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
 else()
    option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
 endif()
@ -82,9 +84,6 @@ endif()

 # dependencies

-set(CMAKE_C_STANDARD   11)
-set(CMAKE_CXX_STANDARD 11)
-
 find_package(Threads REQUIRED)

 # on APPLE - include Accelerate framework
@ -131,6 +130,7 @@ if (WHISPER_ALL_WARNINGS)
            -Wcast-qual                     \
            -Wstrict-prototypes             \
            -Wpointer-arith                 \
+            -Wno-unused-function            \
        ")
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
            -Wall                           \
@ -157,6 +157,7 @@ else()
    if (MSVC)
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
    else()
        if (EMSCRIPTEN)
            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
@ -168,7 +169,10 @@ else()
            if(NOT WHISPER_NO_AVX2)
                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
            endif()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma -mf16c")
+            if(NOT WHISPER_NO_FMA)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
+            endif()
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
        endif()
    endif()
 endif()
@ -190,6 +194,8 @@ add_library(${TARGET}
    whisper.cpp
    )

+include(DefaultTargetOptions)
+
 target_include_directories(${TARGET} PUBLIC
    .
    )
@ -223,6 +229,7 @@ target_compile_definitions(${TARGET} PUBLIC
 install(TARGETS ${TARGET}
    LIBRARY DESTINATION lib
    ARCHIVE DESTINATION lib/static
+    RUNTIME DESTINATION bin
    )

 #
--- a/29
+++ b/29
@ -10,6 +10,9 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif

+CCV := $(shell $(CC) --version | head -n 1)
+CXXV := $(shell $(CXX) --version | head -n 1)
+
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@ -53,10 +56,13 @@ endif
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),x86_64)
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 	ifeq ($(UNAME_S),Darwin)
-		CFLAGS += -mfma -mf16c
+		CFLAGS += -mf16c
 		AVX1_M := $(shell sysctl machdep.cpu.features)
+		ifneq (,$(findstring FMA,$(AVX1_M)))
+			CFLAGS += -mfma
+		endif
 		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
 			CFLAGS += -mavx
 		endif
@ -81,6 +87,10 @@ ifeq ($(UNAME_M),x86_64)
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
 		endif
+		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
+		ifneq (,$(findstring sse3,$(SSE3_M)))
+			CFLAGS += -msse3
+		endif
 	else ifeq ($(UNAME_S),Haiku)
 		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
 		ifneq (,$(findstring avx,$(AVX1_M)))
@ -141,6 +151,21 @@ ifneq ($(filter armv8%,$(UNAME_M)),)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif

+#
+# Print build information
+#
+
+$(info I whisper.cpp build info: )
+$(info I UNAME_S:  $(UNAME_S))
+$(info I UNAME_P:  $(UNAME_P))
+$(info I UNAME_M:  $(UNAME_M))
+$(info I CFLAGS:   $(CFLAGS))
+$(info I CXXFLAGS: $(CXXFLAGS))
+$(info I LDFLAGS:  $(LDFLAGS))
+$(info I CC:       $(CCV))
+$(info I CXX:      $(CXXV))
+$(info )
+
 default: main

 #
--- a/README.md
+++ b/README.md
@ -11,6 +11,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - Plain C/C++ implementation without dependencies
 - Apple silicon first-class citizen - optimized via Arm Neon and Accelerate framework
 - AVX intrinsics support for x86 architectures
+- VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - Low memory usage (Flash Attention + Flash Forward)
 - Zero memory allocations at runtime
@ -70,7 +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
 ```

 ---
@ -88,27 +89,36 @@ 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
+  -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
+  -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 ...
@ -211,17 +221,7 @@ make large
 ## 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

@ -306,6 +306,7 @@ 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
 ```

--- a/bindings/go/.gitignore
+++ b/bindings/go/.gitignore
@ -1,3 +1,2 @@
 build
 models
-go.sum
--- a/bindings/go/Makefile
+++ b/bindings/go/Makefile
@ -1,28 +1,27 @@
-CMAKE := $(shell which cmake)
-BUILD_DIR := "build"
-MODELS_DIR := "models"
+BUILD_DIR := build
+MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
-C_INCLUDE_PATH := "../.."
+INCLUDE_PATH := $(abspath ../..)
+LIBRARY_PATH := $(abspath ../..)

 all: clean whisper examples

 whisper: mkdir
 	@echo Build whisper
-	@${CMAKE} -S ../.. -B ${BUILD_DIR} -D BUILD_SHARED_LIBS=off -D WHISPER_NO_AVX2=on
-	@${CMAKE} --build ${BUILD_DIR} --target whisper
+	@${MAKE} -C ../.. libwhisper.a

 test: model-small whisper modtidy
-	@go test -v .
-	@go test -v ./pkg/whisper/...
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v .
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v ./pkg/whisper/...

 examples: $(EXAMPLES_DIR)

 model-small: mkdir examples/go-model-download
-	@${BUILD_DIR}/go-model-download -out models small.en
+	@${BUILD_DIR}/go-model-download -out models ggml-small.en.bin

 $(EXAMPLES_DIR): mkdir whisper modtidy
 	@echo Build example $(notdir $@)
-	@go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
+	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@

 mkdir:
 	@echo Mkdir ${BUILD_DIR}
--- a/bindings/go/README.md
+++ b/bindings/go/README.md
@ -74,4 +74,27 @@ And you can then test a model against samples with the following command:
 ./build/go-whisper -model models/ggml-tiny.en.bin samples/jfk.wav 
 ```

+## Using the bindings
+
+To use the bindings in your own software,
+
+  1. Import `github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper` (or `github.com/ggerganov/whisper.cpp/bindings/go` into your package;
+  2. Compile `libwhisper.a` (you can use `make whisper` in the `bindings/go` directory);
+  3. Link your go binary against whisper by setting the environment variables `C_INCLUDE_PATH` and `LIBRARY_PATH`
+     to point to the `whisper.h` file directory and `libwhisper.a` file directory respectively.
+
+Look at the `Makefile` in the `bindings/go` directory for an example.
+
+The API Documentation:
+
+  * https://pkg.go.dev/github.com/ggerganov/whisper.cpp/bindings/go
+  * https://pkg.go.dev/github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper
+
+Getting help:
+
+  * Follow the discussion for the go bindings [here](https://github.com/ggerganov/whisper.cpp/discussions/312)
+
+## License
+
+The license for the Go bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.

--- a/bindings/go/examples/go-model-download/main.go
+++ b/bindings/go/examples/go-model-download/main.go
@ -17,15 +17,14 @@ import (
 // CONSTANTS

 const (
-	srcUrl        = "https://huggingface.co/"                           // The location of the models
-	srcPathPrefix = "/datasets/ggerganov/whisper.cpp/resolve/main/ggml" // Filename prefix
-	srcExt        = ".bin"                                              // Filename extension
-	bufSize       = 1024 * 64                                           // Size of the buffer used for downloading the model
+	srcUrl  = "https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main" // The location of the models
+	srcExt  = ".bin"                                                               // Filename extension
+	bufSize = 1024 * 64                                                            // Size of the buffer used for downloading the model
 )

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

 var (
@ -123,11 +122,14 @@ func GetModels() []string {

 // URLForModel returns the URL for the given model on huggingface.co
 func URLForModel(model string) (string, error) {
+	if filepath.Ext(model) != srcExt {
+		model += srcExt
+	}
 	url, err := url.Parse(srcUrl)
 	if err != nil {
 		return "", err
 	} else {
-		url.Path = srcPathPrefix + "-" + model + srcExt
+		url.Path = filepath.Join(url.Path, model)
 	}
 	return url.String(), nil
 }
--- a/bindings/go/examples/go-whisper/color.go
+++ b/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
+}
--- a/bindings/go/examples/go-whisper/flags.go
+++ b/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)")
 }
--- a/bindings/go/examples/go-whisper/main.go
+++ b/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
 		}
--- a/bindings/go/examples/go-whisper/process.go
+++ b/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 +20,20 @@ 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
+	}
+
 	// 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 +45,83 @@ 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)
 	if err := context.Process(data, cb); err != nil {
 		return err
 	}

 	// 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)
 }
--- a/bindings/go/go.sum
+++ b/bindings/go/go.sum
@ -0,0 +1,23 @@
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/go-audio/audio v1.0.0 h1:zS9vebldgbQqktK4H0lUqWrG8P0NxCJVqcj7ZpNnwd4=
+github.com/go-audio/audio v1.0.0/go.mod h1:6uAu0+H2lHkwdGsAY+j2wHPNPpPoeg5AaEFh9FlA+Zs=
+github.com/go-audio/riff v1.0.0 h1:d8iCGbDvox9BfLagY94fBynxSPHO80LmZCaOsmKxokA=
+github.com/go-audio/riff v1.0.0/go.mod h1:l3cQwc85y79NQFCRB7TiPoNiaijp6q8Z0Uv38rVG498=
+github.com/go-audio/wav v1.1.0 h1:jQgLtbqBzY7G+BM8fXF7AHUk1uHUviWS4X39d5rsL2g=
+github.com/go-audio/wav v1.1.0/go.mod h1:mpe9qfwbScEbkd8uybLuIpTgHyrISw/OTuvjUW2iGtE=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
+github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=
+github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
+github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=
+github.com/stretchr/testify v1.8.1 h1:w7B6lhMri9wdJUVmEZPGGhZzrYTPvgJArz7wNPgYKsk=
+github.com/stretchr/testify v1.8.1/go.mod h1:w2LPCIKwWwSfY2zedu0+kehJoqGctiVI29o6fzry7u4=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
--- a/bindings/go/params.go
+++ b/bindings/go/params.go
@ -1,8 +1,5 @@
 package whisper

-// This file defines the whisper_token, whisper_token_data and whisper_full_params
-// structures, which are used by the whisper_full() function.
-
 import (
 	"fmt"
 )
@ -50,6 +47,7 @@ func (p *Params) SetSpeedup(v bool) {
 	p.speed_up = toBool(v)
 }

+// Set language id
 func (p *Params) SetLanguage(lang int) error {
 	str := C.whisper_lang_str(C.int(lang))
 	if str == nil {
@ -60,6 +58,7 @@ func (p *Params) SetLanguage(lang int) error {
 	return nil
 }

+// Get language id
 func (p *Params) Language() int {
 	if p.language == nil {
 		return -1
@ -67,18 +66,41 @@ func (p *Params) Language() int {
 	return int(C.whisper_lang_id(p.language))
 }

+// 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

--- a/bindings/go/pkg/whisper/consts.go
+++ b/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")
 )

 ///////////////////////////////////////////////////////////////////////////////
--- a/bindings/go/pkg/whisper/context.go
+++ b/bindings/go/pkg/whisper/context.go
@ -24,7 +24,7 @@ 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
@ -41,6 +41,9 @@ func (context *context) SetLanguage(lang string) error {
 	if context.model.ctx == nil {
 		return ErrInternalAppError
 	}
+	if !context.model.IsMultilingual() {
+		return ErrModelNotMultilingual
+	}
 	if id := context.model.ctx.Whisper_lang_id(lang); id < 0 {
 		return ErrUnsupportedLanguage
 	} else if err := context.params.SetLanguage(id); err != nil {
@ -50,16 +53,60 @@ 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 {
 	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))
+}
+
 // Process new sample data and return any errors
 func (context *context) Process(data []float32, cb SegmentCallback) error {
 	if context.model.ctx == nil {
@ -119,6 +166,65 @@ 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

--- a/bindings/go/pkg/whisper/interface.go
+++ b/bindings/go/pkg/whisper/interface.go
@ -20,6 +20,9 @@ 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
 }
@ -27,8 +30,18 @@ type Model interface {
 // Context is the speach recognition context.
 type Context interface {
 	SetLanguage(string) error // Set the language to use for speech recognition.
+	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 +51,15 @@ 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
 }

 // Segment is the text result of a speech recognition.
--- a/bindings/go/pkg/whisper/model.go
+++ b/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)
 }
--- a/bindings/go/whisper.go
+++ b/bindings/go/whisper.go
@ -9,8 +9,7 @@ import (
 // CGO

 /*
-#cgo CFLAGS: -I${SRCDIR}/../..
-#cgo LDFLAGS: -L${SRCDIR}/build -lwhisper -lm -lstdc++
+#cgo LDFLAGS: -lwhisper -lm -lstdc++
 #cgo darwin LDFLAGS: -framework Accelerate
 #include <whisper.h>
 #include <stdlib.h>
@ -92,7 +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
@ -171,6 +170,10 @@ func (ctx *Context) Whisper_tokenize(text string, tokens []Token) (int, error) {
 }

 // Return the id of the specified language, returns -1 if not found
+// Examples:
+//
+//	"de" -> 2
+//	"german" -> 2
 func (ctx *Context) Whisper_lang_id(lang string) int {
 	return int(C.whisper_lang_id(C.CString(lang)))
 }
@ -211,6 +214,10 @@ func (ctx *Context) Whisper_n_text_ctx() int {
 	return int(C.whisper_n_text_ctx((*C.struct_whisper_context)(ctx)))
 }

+func (ctx *Context) Whisper_n_audio_ctx() int {
+	return int(C.whisper_n_audio_ctx((*C.struct_whisper_context)(ctx)))
+}
+
 func (ctx *Context) Whisper_is_multilingual() int {
 	return int(C.whisper_is_multilingual((*C.struct_whisper_context)(ctx)))
 }
--- a/bindings/go/whisper_test.go
+++ b/bindings/go/whisper_test.go
@ -50,7 +50,10 @@ func Test_Whisper_001(t *testing.T) {
 	ctx := whisper.Whisper_init(ModelPath)
 	assert.NotNil(ctx)
 	defer ctx.Whisper_free()
-	assert.NoError(ctx.Whisper_full(ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY), buf.AsFloat32Buffer().Data, nil, nil))
+	params := ctx.Whisper_full_default_params(whisper.SAMPLING_GREEDY)
+	data := buf.AsFloat32Buffer().Data
+	err = ctx.Whisper_full(params, data, nil, nil)
+	assert.NoError(err)

 	// Print out tokens
 	num_segments := ctx.Whisper_full_n_segments()
--- a/bindings/ios
+++ b/bindings/ios
--- a/bindings/javascript/emscripten.cpp
+++ b/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 {
--- a/bindings/javascript/whisper.js
+++ b/bindings/javascript/whisper.js
--- a/cmake/DefaultTargetOptions.cmake
+++ b/cmake/DefaultTargetOptions.cmake
@ -0,0 +1,17 @@
+# Set the default compile features and properties for a target.
+
+if (NOT TARGET)
+    message(FATAL_ERROR "TARGET not set before including DefaultTargetOptions")
+endif()
+
+target_compile_features(${TARGET}
+    PRIVATE
+        cxx_std_11
+    )
+
+set_target_properties(${TARGET}
+    PROPERTIES
+        EXPORT_COMPILE_COMMANDS ON
+        RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin"
+        INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib"
+)
--- a/examples/bench.wasm/CMakeLists.txt
+++ b/examples/bench.wasm/CMakeLists.txt
@ -8,6 +8,8 @@ add_executable(${TARGET}
    emscripten.cpp
    )

+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE
    whisper
    )
--- a/examples/bench.wasm/emscripten.cpp
+++ b/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();
--- a/examples/bench/CMakeLists.txt
+++ b/examples/bench/CMakeLists.txt
@ -1,3 +1,6 @@
 set(TARGET bench)
 add_executable(${TARGET} bench.cpp)
+
+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
--- a/examples/bench/bench.cpp
+++ b/examples/bench/bench.cpp
@ -53,7 +53,7 @@ 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());

    {
        fprintf(stderr, "\n");
--- a/examples/command.wasm/CMakeLists.txt
+++ b/examples/command.wasm/CMakeLists.txt
@ -8,6 +8,8 @@ add_executable(${TARGET}
    emscripten.cpp
    )

+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE
    whisper
    )
--- a/examples/command.wasm/emscripten.cpp
+++ b/examples/command.wasm/emscripten.cpp
@ -324,7 +324,7 @@ EMSCRIPTEN_BINDINGS(command) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        for (size_t i = 0; i < g_contexts.size(); ++i) {
            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init(path_model.c_str());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                if (g_contexts[i] != nullptr) {
                    g_running = true;
                    if (g_worker.joinable()) {
--- a/examples/command/CMakeLists.txt
+++ b/examples/command/CMakeLists.txt
@ -2,6 +2,9 @@ if (WHISPER_SUPPORT_SDL2)
    # command
    set(TARGET command)
    add_executable(${TARGET} command.cpp)
+
+    include(DefaultTargetOptions)
+
    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 endif ()
--- a/examples/command/README.md
+++ b/examples/command/README.md
@ -9,7 +9,19 @@ More info is available in [issue #171](https://github.com/ggerganov/whisper.cpp/

 # On Raspberry Pi, use tiny or base models + "-ac 768" for better performance
 ./command -m ./models/ggml-tiny.en.bin -ac 768 -t 3 -c 0
+```

+https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
+
+Web version: [examples/command.wasm](/examples/command.wasm)
+
+## Guided mode
+
+"Guided mode" allows you to specify a list of commands (i.e. strings) and the transcription will be guided to classify your command into one from the list. This can be useful in situations where a device is listening only for a small subset of commands.
+
+Initial tests show that this approach might be extremely efficient in terms of performance, since it integrates very well with the "partial Encoder" idea from #137.
+
+```bash
 # Run in guided mode, the list of allowed commands is in commands.txt
 ./command -m ./models/ggml-base.en.bin -cmd ./examples/command/commands.txt

@ -17,9 +29,8 @@ More info is available in [issue #171](https://github.com/ggerganov/whisper.cpp/
 ./command -m ./models/ggml-tiny.en.bin -cmd ./examples/command/commands.txt -ac 128 -t 3 -c 0
 ```

-https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
+https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9b8b-aeeb76bee969.mp4

-Web version: [examples/command.wasm](/examples/command.wasm)

 ## Building

--- a/examples/command/command.cpp
+++ b/examples/command/command.cpp
@ -11,6 +11,7 @@
 #include <SDL.h>
 #include <SDL_audio.h>

+#include <sstream>
 #include <cassert>
 #include <cstdio>
 #include <fstream>
@ -25,7 +26,7 @@
 struct whisper_params {
    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
    int32_t prompt_ms  = 5000;
-    int32_t command_ms = 4000;
+    int32_t command_ms = 8000;
    int32_t capture_id = -1;
    int32_t max_tokens = 32;
    int32_t audio_ctx  = 0;
@ -43,6 +44,7 @@ struct whisper_params {
    std::string model     = "models/ggml-base.en.bin";
    std::string fname_out;
    std::string commands;
+    std::string prompt;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@ -71,6 +73,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
        else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
+        else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -103,6 +106,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
    fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
+    fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
    fprintf(stderr, "\n");
 }

@ -510,6 +514,433 @@ std::vector<std::string> read_allowed_commands(const std::string & fname) {
    return allowed_commands;
 }

+std::vector<std::string> get_words(const std::string &txt) {
+    std::vector<std::string> words;
+
+    std::istringstream iss(txt);
+    std::string word;
+    while (iss >> word) {
+        words.push_back(word);
+    }
+
+    return words;
+}
+
+// returns true if no exit event was received
+bool process_sdl_events() {
+    SDL_Event event;
+    while (SDL_PollEvent(&event)) {
+        switch (event.type) {
+            case SDL_QUIT:
+                {
+                    return false;
+                } break;
+            default:
+                break;
+        }
+    }
+
+    return true;
+}
+
+// command-list mode
+// guide the transcription to match the most likely command from a provided list
+int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: guided mode\n", __func__);
+
+    std::vector<std::string> allowed_commands = read_allowed_commands(params.commands);
+
+    if (allowed_commands.empty()) {
+        fprintf(stderr, "%s: error: failed to read allowed commands from '%s'\n", __func__, params.commands.c_str());
+        return 2;
+    }
+
+    int max_len = 0;
+
+    std::vector<std::vector<whisper_token>> allowed_tokens;
+
+    for (const auto & cmd : allowed_commands) {
+        whisper_token tokens[1024];
+        allowed_tokens.emplace_back();
+
+        for (int l = 0; l < (int) cmd.size(); ++l) {
+            // NOTE: very important to add the whitespace !
+            //       the reason is that the first decoded token starts with a whitespace too!
+            std::string ss = std::string(" ") + cmd.substr(0, l + 1);
+
+            const int n = whisper_tokenize(ctx, ss.c_str(), tokens, 1024);
+            if (n < 0) {
+                fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, cmd.c_str());
+                return 3;
+            }
+
+            if (n == 1) {
+                allowed_tokens.back().push_back(tokens[0]);
+            }
+        }
+
+        max_len = std::max(max_len, (int) cmd.size());
+    }
+
+    fprintf(stderr, "%s: allowed commands [ tokens ]:\n", __func__);
+    fprintf(stderr, "\n");
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        fprintf(stderr, "  - \033[1m%-*s\033[0m = [", max_len, allowed_commands[i].c_str());
+        for (const auto & token : allowed_tokens[i]) {
+            fprintf(stderr, " %5d", token);
+        }
+        fprintf(stderr, " ]\n");
+    }
+
+    std::string  k_prompt = "select one from the available words: ";
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        if (i > 0) {
+            k_prompt += ", ";
+        }
+        k_prompt += allowed_commands[i];
+    }
+    k_prompt += ". selected word: ";
+
+    // tokenize prompt
+    std::vector<whisper_token> k_tokens;
+    {
+        k_tokens.resize(1024);
+        const int n = whisper_tokenize(ctx, k_prompt.c_str(), k_tokens.data(), 1024);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to tokenize prompt '%s'\n", __func__, k_prompt.c_str());
+            return 4;
+        }
+        k_tokens.resize(n);
+    }
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: prompt: '%s'\n", __func__, k_prompt.c_str());
+    fprintf(stderr, "%s: tokens: [", __func__);
+    for (const auto & token : k_tokens) {
+        fprintf(stderr, " %d", token);
+    }
+    fprintf(stderr, " ]\n");
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: listening for a command ...\n", __func__);
+    fprintf(stderr, "\n");
+
+    bool is_running  = true;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = process_sdl_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        audio.get(2000, pcmf32_cur);
+
+        if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+            const auto t_start = std::chrono::high_resolution_clock::now();
+
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.print_progress   = false;
+            wparams.print_special    = params.print_special;
+            wparams.print_realtime   = false;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.translate        = params.translate;
+            wparams.no_context       = true;
+            wparams.single_segment   = true;
+            wparams.max_tokens       = 1;
+            wparams.language         = params.language.c_str();
+            wparams.n_threads        = params.n_threads;
+
+            wparams.audio_ctx        = params.audio_ctx;
+            wparams.speed_up         = params.speed_up;
+
+            wparams.prompt_tokens    = k_tokens.data();
+            wparams.prompt_n_tokens  = k_tokens.size();
+
+            // run the transformer and a single decoding pass
+            if (whisper_full(ctx, wparams, pcmf32_cur.data(), pcmf32_cur.size()) != 0) {
+                fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+                break;
+            }
+
+            // estimate command probability
+            // NOTE: not optimal
+            {
+                const auto * logits = whisper_get_logits(ctx);
+
+                std::vector<float> probs(whisper_n_vocab(ctx), 0.0f);
+
+                // compute probs from logits via softmax
+                {
+                    float max = -1e9;
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        max = std::max(max, logits[i]);
+                    }
+
+                    float sum = 0.0f;
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        probs[i] = expf(logits[i] - max);
+                        sum += probs[i];
+                    }
+
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        probs[i] /= sum;
+                    }
+                }
+
+                std::vector<std::pair<float, int>> probs_id;
+
+                double psum = 0.0;
+                for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+                    probs_id.emplace_back(probs[allowed_tokens[i][0]], i);
+                    for (int j = 1; j < (int) allowed_tokens[i].size(); ++j) {
+                        probs_id.back().first += probs[allowed_tokens[i][j]];
+                    }
+                    probs_id.back().first /= allowed_tokens[i].size();
+                    psum += probs_id.back().first;
+                }
+
+                // normalize
+                for (auto & p : probs_id) {
+                    p.first /= psum;
+                }
+
+                // sort descending
+                {
+                    using pair_type = decltype(probs_id)::value_type;
+                    std::sort(probs_id.begin(), probs_id.end(), [](const pair_type & a, const pair_type & b) {
+                        return a.first > b.first;
+                    });
+                }
+
+                // print the commands and the respective probabilities
+                {
+                    fprintf(stdout, "\n");
+                    for (const auto & cmd : probs_id) {
+                        fprintf(stdout, "%s: %s%-*s%s = %f | ", __func__, "\033[1m", max_len, allowed_commands[cmd.second].c_str(), "\033[0m", cmd.first);
+                        for (int token : allowed_tokens[cmd.second]) {
+                            fprintf(stdout, "'%4s' %f ", whisper_token_to_str(ctx, token), probs[token]);
+                        }
+                        fprintf(stdout, "\n");
+                    }
+                }
+
+                // best command
+                {
+                    const auto t_end = std::chrono::high_resolution_clock::now();
+
+                    const float prob = probs_id[0].first;
+                    const int index = probs_id[0].second;
+
+                    fprintf(stdout, "\n");
+                    fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
+                            "\033[1m", allowed_commands[index].c_str(), "\033[0m", prob,
+                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
+                    fprintf(stdout, "\n");
+                }
+            }
+
+            audio.clear();
+        }
+    }
+
+    return 0;
+}
+
+// always-prompt mode
+// transcribe the voice into text after valid prompt
+int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+    bool is_running = true;
+    bool ask_prompt = true;
+
+    float prob = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+
+    const std::string k_prompt = params.prompt;
+
+    const int k_prompt_length = get_words(k_prompt).size();
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: always-prompt mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = process_sdl_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: The prompt is: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                int64_t t_ms = 0;
+
+                // detect the commands
+                audio.get(params.command_ms, pcmf32_cur);
+
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+
+                const auto words = get_words(txt);
+
+                std::string prompt;
+                std::string command;
+
+                for (int i = 0; i < (int) words.size(); ++i) {
+                    if (i < k_prompt_length) {
+                        prompt += words[i] + " ";
+                    } else {
+                        command += words[i] + " ";
+                    }
+                }
+
+                const float sim = similarity(prompt, k_prompt);
+
+                //debug
+                //fprintf(stdout, "command size: %i\n", command_length);
+
+                if ((sim > 0.7f) && (command.size() > 0)) {
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                }
+
+                fprintf(stdout, "\n");
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
+// general-purpose mode
+// freely transcribe the voice into text
+int process_general_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    bool is_running  = true;
+    bool have_prompt = false;
+    bool ask_prompt  = true;
+
+    float prob0 = 0.0f;
+    float prob  = 0.0f;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string k_prompt = "Ok Whisper, start listening for commands.";
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: general-purpose mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = process_sdl_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                int64_t t_ms = 0;
+
+                if (!have_prompt) {
+                    // wait for activation phrase
+                    audio.get(params.prompt_ms, pcmf32_cur);
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
+
+                    const float sim = similarity(txt, k_prompt);
+
+                    if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                        fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                        ask_prompt = true;
+                    } else {
+                        fprintf(stdout, "\n");
+                        fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                        fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                        fprintf(stdout, "\n");
+
+                        // save the audio for the prompt
+                        pcmf32_prompt = pcmf32_cur;
+                        have_prompt = true;
+                    }
+                } else {
+                    // we have heard the activation phrase, now detect the commands
+                    audio.get(params.command_ms, pcmf32_cur);
+
+                    // prepend the prompt audio
+                    pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
+
+                    prob = 100.0f*(prob - prob0);
+
+                    //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                    // find the prompt in the text
+                    float best_sim = 0.0f;
+                    size_t best_len = 0;
+                    for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        const auto prompt = txt.substr(0, n);
+
+                        const float sim = similarity(prompt, k_prompt);
+
+                        //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                        if (sim > best_sim) {
+                            best_sim = sim;
+                            best_len = n;
+                        }
+                    }
+
+                    const std::string command = ::trim(txt.substr(best_len));
+
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    fprintf(stdout, "\n");
+                }
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
 int main(int argc, char ** argv) {
    whisper_params params;

@ -525,7 +956,7 @@ 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());

    // print some info about the processing
    {
@ -561,300 +992,14 @@ int main(int argc, char ** argv) {
    std::this_thread::sleep_for(std::chrono::milliseconds(1000));
    audio.clear();

-    int max_len = 0;
-
-    bool is_running  = true;
-    bool have_prompt = false;
-    bool ask_prompt  = true;
-
-    float prob0 = 0.0f;
-    float prob  = 0.0f;
-
-    std::vector<float> pcmf32_cur;
-    std::vector<float> pcmf32_prompt;
-
-    std::vector<std::string> allowed_commands;
-    std::vector<std::vector<whisper_token>> allowed_tokens;
-
-    std::string k_prompt;
-    std::vector<whisper_token> k_tokens;
+    int  ret_val = 0;

    if (!params.commands.empty()) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: guided mode\n", __func__);
-
-        allowed_commands = read_allowed_commands(params.commands);
-
-        if (allowed_commands.empty()) {
-            fprintf(stderr, "%s: error: failed to read allowed commands from '%s'\n", __func__, params.commands.c_str());
-            return 2;
-        }
-
-        for (const auto & cmd : allowed_commands) {
-            whisper_token tokens[1024];
-            allowed_tokens.emplace_back();
-
-            for (int l = 0; l < (int) cmd.size(); ++l) {
-                // NOTE: very important to add the whitespace !
-                //       the reason is that the first decoded token starts with a whitespace too!
-                std::string ss = std::string(" ") + cmd.substr(0, l + 1);
-
-                const int n = whisper_tokenize(ctx, ss.c_str(), tokens, 1024);
-                if (n < 0) {
-                    fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, cmd.c_str());
-                    return 3;
-                }
-
-                if (n == 1) {
-                    allowed_tokens.back().push_back(tokens[0]);
-                }
-            }
-
-            max_len = std::max(max_len, (int) cmd.size());
-        }
-
-        fprintf(stderr, "%s: allowed commands [ tokens ]:\n", __func__);
-        fprintf(stderr, "\n");
-        for (int i = 0; i < (int) allowed_commands.size(); ++i) {
-            fprintf(stderr, "  - \033[1m%-*s\033[0m = [", max_len, allowed_commands[i].c_str());
-            for (const auto & token : allowed_tokens[i]) {
-                fprintf(stderr, " %5d", token);
-            }
-            fprintf(stderr, " ]\n");
-        }
-
-        k_prompt = "select one from the available words: ";
-        for (int i = 0; i < (int) allowed_commands.size(); ++i) {
-            if (i > 0) {
-                k_prompt += ", ";
-            }
-            k_prompt += allowed_commands[i];
-        }
-        k_prompt += ". selected word: ";
-
-        // tokenize prompt
-        {
-            k_tokens.resize(1024);
-            const int n = whisper_tokenize(ctx, k_prompt.c_str(), k_tokens.data(), 1024);
-            if (n < 0) {
-                fprintf(stderr, "%s: error: failed to tokenize prompt '%s'\n", __func__, k_prompt.c_str());
-                return 4;
-            }
-            k_tokens.resize(n);
-        }
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, k_prompt.c_str());
-        fprintf(stderr, "%s: tokens: [", __func__);
-        for (const auto & token : k_tokens) {
-            fprintf(stderr, " %d", token);
-        }
-        fprintf(stderr, " ]\n");
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: listening for a command ...\n", __func__);
-        fprintf(stderr, "\n");
-
+        ret_val = process_command_list(ctx, audio, params);
+    } else if (!params.prompt.empty()) {
+        ret_val = always_prompt_transcription(ctx, audio, params);
    } else {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: general-purpose mode\n", __func__);
-
-        k_prompt = "Ok Whisper, start listening for commands.";
-    }
-
-    // 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;
-                }
-            }
-
-            if (!is_running) {
-                break;
-            }
-        }
-
-        // delay
-        std::this_thread::sleep_for(std::chrono::milliseconds(100));
-
-        if (allowed_commands.empty()) {
-            // general-purpose mode
-            // freely transcribe the voice into text
-
-            if (ask_prompt) {
-                fprintf(stdout, "\n");
-                fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
-                fprintf(stdout, "\n");
-
-                ask_prompt = false;
-            }
-
-            {
-                int64_t t_ms = 0;
-
-                audio.get(2000, pcmf32_cur);
-
-                if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
-                    fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
-
-                    if (!have_prompt) {
-                        // wait for activation phrase
-                        audio.get(params.prompt_ms, pcmf32_cur);
-
-                        const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob0, t_ms));
-
-                        fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms);
-
-                        const float sim = similarity(txt, k_prompt);
-
-                        if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
-                            fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
-                            ask_prompt = true;
-                        } else {
-                            fprintf(stdout, "\n");
-                            fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
-                            fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
-                            fprintf(stdout, "\n");
-
-                            // save the audio for the prompt
-                            pcmf32_prompt = pcmf32_cur;
-                            have_prompt = true;
-                        }
-                    } else {
-                        // we have heard the activation phrase, now detect the commands
-                        audio.get(params.command_ms, pcmf32_cur);
-
-                        // prepend the prompt audio
-                        pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
-
-                        const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, prob, t_ms));
-
-                        prob = 100.0f*(prob - prob0);
-
-                        //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
-
-                        // find the prompt in the text
-                        float best_sim = 0.0f;
-                        size_t best_len = 0;
-                        for (int n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
-                            const auto prompt = txt.substr(0, n);
-
-                            const float sim = similarity(prompt, k_prompt);
-
-                            //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
-
-                            if (sim > best_sim) {
-                                best_sim = sim;
-                                best_len = n;
-                            }
-                        }
-
-                        const std::string command = ::trim(txt.substr(best_len));
-
-                        fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
-                        fprintf(stdout, "\n");
-                    }
-
-                    audio.clear();
-                }
-            }
-        } else {
-            // command-list mode
-            // guide the transcription to match the most likely command from a provided list
-
-            audio.get(2000, pcmf32_cur);
-
-            if (vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
-                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
-
-                const auto t_start = std::chrono::high_resolution_clock::now();
-
-                whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-
-                wparams.print_progress   = false;
-                wparams.print_special    = params.print_special;
-                wparams.print_realtime   = false;
-                wparams.print_timestamps = !params.no_timestamps;
-                wparams.translate        = params.translate;
-                wparams.no_context       = true;
-                wparams.single_segment   = true;
-                wparams.max_tokens       = 1;
-                wparams.language         = params.language.c_str();
-                wparams.n_threads        = params.n_threads;
-
-                wparams.audio_ctx        = params.audio_ctx;
-                wparams.speed_up         = params.speed_up;
-
-                wparams.prompt_tokens    = k_tokens.data();
-                wparams.prompt_n_tokens  = k_tokens.size();
-
-                // run the transformer and a single decoding pass
-                if (whisper_full(ctx, wparams, pcmf32_cur.data(), pcmf32_cur.size()) != 0) {
-                    fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
-                    break;
-                }
-
-                const auto * probs = whisper_get_probs(ctx);
-                std::vector<std::pair<float, int>> probs_id;
-
-                double psum = 0.0;
-                for (int i = 0; i < (int) allowed_commands.size(); ++i) {
-                    probs_id.emplace_back(probs[allowed_tokens[i][0]], i);
-                    for (int j = 1; j < (int) allowed_tokens[i].size(); ++j) {
-                        probs_id.back().first += probs[allowed_tokens[i][j]];
-                    }
-                    probs_id.back().first /= allowed_tokens[i].size();
-                    psum += probs_id.back().first;
-                }
-
-                // normalize
-                for (auto & p : probs_id) {
-                    p.first /= psum;
-                }
-
-                // sort descending
-                {
-                    using pair_type = decltype(probs_id)::value_type;
-                    std::sort(probs_id.begin(), probs_id.end(), [](const pair_type & a, const pair_type & b) {
-                        return a.first > b.first;
-                    });
-                }
-
-                // print the commands and the respective probabilities
-                {
-                    fprintf(stdout, "\n");
-                    for (const auto & cmd : probs_id) {
-                        fprintf(stdout, "%s: %s%-*s%s = %f | ", __func__, "\033[1m", max_len, allowed_commands[cmd.second].c_str(), "\033[0m", cmd.first);
-                        for (int i = 0; i < (int) allowed_tokens[cmd.second].size(); ++i) {
-                            fprintf(stdout, "'%4s' %f ", whisper_token_to_str(ctx, allowed_tokens[cmd.second][i]), probs[allowed_tokens[cmd.second][i]]);
-                        }
-                        fprintf(stdout, "\n");
-                    }
-                }
-
-                // best command
-                {
-                    const auto t_end = std::chrono::high_resolution_clock::now();
-
-                    fprintf(stdout, "\n");
-                    fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
-                            "\033[1m", allowed_commands[probs_id[0].second].c_str(), "\033[0m", probs_id[0].first,
-                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
-                    fprintf(stdout, "\n");
-                }
-
-                audio.clear();
-            }
-        }
+        ret_val = process_general_transcription(ctx, audio, params);
    }

    audio.pause();
@ -862,5 +1007,5 @@ int main(int argc, char ** argv) {
    whisper_print_timings(ctx);
    whisper_free(ctx);

-    return 0;
+    return ret_val;
 }
--- a/examples/main/CMakeLists.txt
+++ b/examples/main/CMakeLists.txt
@ -1,3 +1,6 @@
 set(TARGET main)
 add_executable(${TARGET} main.cpp)
+
+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@ -59,8 +59,12 @@ struct whisper_params {
    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 word_thold    = 0.01f;
+    float entropy_thold = 2.4f;
+    float logprob_thold = -1.0f;

    bool speed_up       = false;
    bool translate      = false;
@ -69,6 +73,7 @@ struct whisper_params {
    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;
@ -103,7 +108,11 @@ 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; }
@ -111,6 +120,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        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 == "-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; }
@ -134,30 +144,35 @@ 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, "  -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, "  -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, "  -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, "\n");
 }

@ -173,90 +188,81 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi

    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) {
-            if (params.print_colors) {
-                for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-                    if (params.print_special == false) {
-                        const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-                        if (id >= whisper_token_eot(ctx)) {
-                            continue;
-                        }
-                    }
-
-                    const char * text = whisper_full_get_token_text(ctx, i, j);
-                    const float  p    = whisper_full_get_token_p   (ctx, i, j);
-
-                    const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));
-
-                    printf("%s%s%s", k_colors[col].c_str(), text, "\033[0m");
-                }
-            } else {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-                printf("%s", text);
-            }
-            fflush(stdout);
-        } else {
-            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-
-            std::string speaker;
-
-            if (params.diarize && pcmf32s.size() == 2) {
-                const int64_t n_samples = pcmf32s[0].size();
-
-                const int64_t is0 = timestamp_to_sample(t0, n_samples);
-                const int64_t is1 = timestamp_to_sample(t1, n_samples);
-
-                double energy0 = 0.0f;
-                double energy1 = 0.0f;
-
-                for (int64_t j = is0; j < is1; j++) {
-                    energy0 += fabs(pcmf32s[0][j]);
-                    energy1 += fabs(pcmf32s[1][j]);
-                }
-
-                if (energy0 > 1.1*energy1) {
-                    speaker = "(speaker 0)";
-                } else if (energy1 > 1.1*energy0) {
-                    speaker = "(speaker 1)";
-                } else {
-                    speaker = "(speaker ?)";
-                }
-
-                //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, %s\n", is0, is1, energy0, energy1, speaker.c_str());
-            }
-
-            if (params.print_colors) {
-                printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
-                for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-                    if (params.print_special == false) {
-                        const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-                        if (id >= whisper_token_eot(ctx)) {
-                            continue;
-                        }
-                    }
-
-                    const char * text = whisper_full_get_token_text(ctx, i, j);
-                    const float  p    = whisper_full_get_token_p   (ctx, i, j);
-
-                    const int col = std::max(0, std::min((int) k_colors.size(), (int) (std::pow(p, 3)*float(k_colors.size()))));
-
-                    printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
-                }
-                printf("\n");
-            } else {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-
-                printf("[%s --> %s]  %s%s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), speaker.c_str(), text);
-            }
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            const int64_t n_samples = pcmf32s[0].size();
+
+            const int64_t is0 = timestamp_to_sample(t0, n_samples);
+            const int64_t is1 = timestamp_to_sample(t1, n_samples);
+
+            double energy0 = 0.0f;
+            double energy1 = 0.0f;
+
+            for (int64_t j = is0; j < is1; j++) {
+                energy0 += fabs(pcmf32s[0][j]);
+                energy1 += fabs(pcmf32s[1][j]);
+            }
+
+            if (energy0 > 1.1*energy1) {
+                speaker = "(speaker 0)";
+            } else if (energy1 > 1.1*energy0) {
+                speaker = "(speaker 1)";
+            } else {
+                speaker = "(speaker ?)";
+            }
+
+            //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, %s\n", is0, is1, energy0, energy1, speaker.c_str());
+        }
+
+        if (params.print_colors) {
+            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
+                if (params.print_special == false) {
+                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
+                    if (id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+                }
+
+                const char * text = whisper_full_get_token_text(ctx, i, j);
+                const float  p    = whisper_full_get_token_p   (ctx, i, j);
+
+                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
+
+                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
+            }
+        } else {
+            const char * text = whisper_full_get_segment_text(ctx, i);
+
+            printf("%s%s", speaker.c_str(), text);
+        }
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+
+        fflush(stdout);
    }
 }

@ -325,6 +331,31 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
    return true;
 }

+bool output_csv(struct whisper_context * ctx, const char * fname) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    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";
+    }
+
+    return true;
+}
+
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
@ -458,7 +489,7 @@ 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");
@ -528,7 +559,7 @@ int main(int argc, char ** argv) {
            }

            if (wav.sampleRate != WHISPER_SAMPLE_RATE) {
-                fprintf(stderr, "%s: WAV file '%s' must be 16 kHz\n", argv[0], fname_inp.c_str());
+                fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", argv[0], fname_inp.c_str(), WHISPER_SAMPLE_RATE/1000);
                return 8;
            }

@ -600,6 +631,8 @@ 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;
@ -613,12 +646,18 @@ int main(int argc, char ** argv) {

            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.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 = -1;
+
+            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 };

@ -674,6 +713,13 @@ int main(int argc, char ** argv) {
                const auto fname_wts = fname_inp + ".wts";
                output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE);
            }
+
+	    // output to CSV file
+            if (params.output_csv) {
+                const auto fname_csv = fname_inp + ".csv";
+                output_csv(ctx, fname_csv.c_str());
+            }
+
        }
    }

--- a/examples/stream.wasm/CMakeLists.txt
+++ b/examples/stream.wasm/CMakeLists.txt
@ -8,6 +8,8 @@ add_executable(${TARGET}
    emscripten.cpp
    )

+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE
    whisper
    )
--- a/examples/stream.wasm/emscripten.cpp
+++ b/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()) {
--- a/examples/stream/CMakeLists.txt
+++ b/examples/stream/CMakeLists.txt
@ -2,6 +2,9 @@ if (WHISPER_SUPPORT_SDL2)
    # stream
    set(TARGET stream)
    add_executable(${TARGET} stream.cpp)
+
+    include(DefaultTargetOptions)
+
    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
    target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 endif ()
--- a/examples/stream/stream.cpp
+++ b/examples/stream/stream.cpp
@ -8,6 +8,7 @@
 #include <SDL.h>
 #include <SDL_audio.h>

+#include <atomic>
 #include <cassert>
 #include <cstdio>
 #include <string>
@ -144,8 +145,8 @@ private:
    int m_len_ms = 0;
    int m_sample_rate = 0;

-    bool       m_running = false;
-    std::mutex m_mutex;
+    std::atomic_bool m_running;
+    std::mutex       m_mutex;

    std::vector<float> m_audio;
    std::vector<float> m_audio_new;
@ -155,6 +156,8 @@ private:

 audio_async::audio_async(int len_ms) {
    m_len_ms = len_ms;
+
+    m_running = false;
 }

 audio_async::~audio_async() {
@ -427,10 +430,10 @@ int main(int argc, char ** argv) {
    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_new_line = params.length_ms / params.step_ms - 1; // number of steps to print new line
-
    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
+
    params.no_timestamps = !use_vad;
    params.no_context    = use_vad;
    params.max_tokens    = 0;
@ -453,10 +456,10 @@ 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;
@ -612,6 +615,9 @@ 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();

--- a/examples/talk.wasm/CMakeLists.txt
+++ b/examples/talk.wasm/CMakeLists.txt
@ -9,6 +9,8 @@ add_executable(${TARGET}
    gpt-2.cpp
    )

+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE
    whisper
    )
@ -31,8 +33,8 @@ set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
    --bind \
    -s USE_PTHREADS=1 \
    -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1600MB \
-    -s TOTAL_MEMORY=1600MB \
+    -s INITIAL_MEMORY=1800MB \
+    -s TOTAL_MEMORY=1800MB \
    -s FORCE_FILESYSTEM=1 \
    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
    ${EXTRA_FLAGS} \
--- a/examples/talk.wasm/README.md
+++ b/examples/talk.wasm/README.md
@ -36,7 +36,7 @@ In order to run this demo efficiently, you need to have the following:
 - Latest Chrome or Firefox browser (Safari is not supported)
 - Run this on a desktop or laptop with modern CPU (a mobile phone will likely not be good enough)
 - Speak phrases that are no longer than 10 seconds - this is the audio context of the AI
- The web-page uses about 1.6GB of RAM
+- The web-page uses about 1.8GB of RAM

 Notice that this demo is using the smallest GPT-2 model, so the generated text responses are not always very good.
 Also, the prompting strategy can likely be improved to achieve better results.
--- a/examples/talk.wasm/emscripten.cpp
+++ b/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()) {
--- a/examples/talk/CMakeLists.txt
+++ b/examples/talk/CMakeLists.txt
@ -8,6 +8,9 @@ if (WHISPER_SUPPORT_SDL2)
    # TODO: this is temporary
    #       need to export ggml symbols for MSVC, but too lazy ..
    add_executable(${TARGET} talk.cpp gpt-2.cpp ../../ggml.c ../../whisper.cpp)
+
+    include(DefaultTargetOptions)
+
    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 endif ()
--- a/examples/talk/talk.cpp
+++ b/examples/talk/talk.cpp
@ -498,7 +498,7 @@ 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

--- a/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
+++ b/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
@ -64,16 +64,22 @@ 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) {
+            val inputstream = application.assets.open("models/" + models[0])
+            whisperContext = WhisperContext.createContextFromInputStream(inputstream)
+            printMessage("Loaded model ${models[0]}.\n")
+        }
+
+        //val firstModel = modelsPath.listFiles()!!.first()
+        //whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
    }

    fun transcribeSample() = viewModelScope.launch {
--- a/examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt
+++ b/examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt
@ -4,6 +4,7 @@ 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 +40,22 @@ 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)
+        }
    }
 }

@ -76,6 +86,7 @@ private class WhisperLib {
        }

        // JNI methods
+        external fun initContextFromInputStream(inputStream: InputStream): Long
        external fun initContext(modelPath: String): Long
        external fun freeContext(contextPtr: Long)
        external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)
--- a/examples/whisper.android/app/src/main/jni/whisper/jni.c
+++ b/examples/whisper.android/app/src/main/jni/whisper/jni.c
@ -2,6 +2,7 @@
 #include <android/log.h>
 #include <stdlib.h>
 #include <sys/sysinfo.h>
+#include <string.h>
 #include "whisper.h"

 #define UNUSED(x) (void)(x)
@ -17,13 +18,86 @@ 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;
+}
+
 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;
 }
--- a/examples/whisper.android/local.properties
+++ b/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
--- a/examples/whisper.objc/whisper.objc/ViewController.m
+++ b/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) {
--- a/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift
+++ b/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 {
--- a/examples/whisper.swiftui/whisper.swiftui.demo/Resources/models/.gitignore
+++ b/examples/whisper.swiftui/whisper.swiftui.demo/Resources/models/.gitignore
--- a/examples/whisper.swiftui/whisper.swiftui.demo/Resources/samples/.gitignore
+++ b/examples/whisper.swiftui/whisper.swiftui.demo/Resources/samples/.gitignore
--- a/examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj
+++ b/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 */ = {
--- a/examples/whisper.wasm/CMakeLists.txt
+++ b/examples/whisper.wasm/CMakeLists.txt
@ -8,6 +8,8 @@ add_executable(${TARGET}
    emscripten.cpp
    )

+include(DefaultTargetOptions)
+
 target_link_libraries(${TARGET} PRIVATE
    whisper
    )
--- a/examples/whisper.wasm/emscripten.cpp
+++ b/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 {
--- a/ggml.c
+++ b/ggml.c
--- a/ggml.h
+++ b/ggml.h
@ -177,13 +177,11 @@ extern "C" {
 #include <stddef.h>
 #include <stdbool.h>

-#define GGML_MAX_DIMS         4
-#define GGML_MAX_NODES        4096
-#define GGML_MAX_PARAMS       16
-#define GGML_MAX_CONTEXTS     64
-#define GGML_MAX_OPT          4
-#define GGML_MAX_THREADS      64
-#define GGML_MAX_THREAD_POOLS 16
+#define GGML_MAX_DIMS     4
+#define GGML_MAX_NODES    4096
+#define GGML_MAX_PARAMS   16
+#define GGML_MAX_CONTEXTS 64
+#define GGML_MAX_OPT      4

 #ifdef __ARM_NEON
 // we use the built-in 16-bit float type
@ -733,6 +731,8 @@ int ggml_cpu_has_f16c(void);
 int ggml_cpu_has_fp16_va(void);
 int ggml_cpu_has_wasm_simd(void);
 int ggml_cpu_has_blas(void);
+int ggml_cpu_has_sse3(void);
+int ggml_cpu_has_vsx(void);

 #ifdef  __cplusplus
 }
--- a/models/convert-h5-to-ggml.py
+++ b/models/convert-h5-to-ggml.py
@ -56,7 +56,7 @@ def bytes_to_unicode():
    The reversible bpe codes work on unicode strings.
    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
    And avoids mapping to whitespace/control characters the bpe code barfs on.
    """
--- a/models/download-ggml-model.cmd
+++ b/models/download-ggml-model.cmd
@ -40,7 +40,7 @@ if exist "ggml-%model%.bin" (
  goto :eof
 )

-PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/datasets/ggerganov/whisper.cpp/raw/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
+PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"

 if %ERRORLEVEL% neq 0 (
  echo Failed to download ggml model %model%
--- a/whisper.cpp
+++ b/whisper.cpp
--- a/whisper.h
+++ b/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);
@ -124,6 +137,7 @@ 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 +145,6 @@ 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
@ -148,7 +154,7 @@ extern "C" {
            struct whisper_context * ctx,
                        const char * text,
                     whisper_token * tokens,
-	                           int   n_max_tokens);
+                               int   n_max_tokens);

    // Largest language id (i.e. number of available languages - 1)
    WHISPER_API int whisper_lang_max_id();
@ -177,10 +183,14 @@ extern "C" {
    WHISPER_API int whisper_n_len          (struct whisper_context * ctx); // mel length
    WHISPER_API int whisper_n_vocab        (struct whisper_context * ctx);
    WHISPER_API int whisper_n_text_ctx     (struct whisper_context * ctx);
+    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);
@ -209,8 +219,8 @@ 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
@ -230,17 +240,17 @@ 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 initial prompt for the decoder (if any)
        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
@ -250,10 +260,11 @@ extern "C" {
        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;
@ -261,19 +272,35 @@ 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
+
+        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;
    };
Author	SHA1	Message	Date
Georgi Gerganov	00ea21668b	whisper : account speed_up flag for short audio (close #405 )	2023-01-15 12:42:15 +02:00
Georgi Gerganov	0b85e8c401	Update README.md	2023-01-15 11:36:20 +02:00
Georgi Gerganov	fafd78945d	bench.wasm : print system info	2023-01-15 11:34:03 +02:00
Georgi Gerganov	8de452c18b	Improve decoding (#291 ) * whisper : prepare infra for new decoding strategies * whisper : apply logit filters and compute logprobs * whisper : add whisper_get_logits() * whisper : separate self and cross attention memory Initial step needed for supporting parallel decoders * whisper : move probs_id buffer to whisper_context * whisper : refactor kv cache into separate struct * whisper : move self-attention kv cache to whisper_decoder * whisper : wip decoding parameters + strategies * whisper : wip decoding parameters + strategies (part 2) * whisper : wip decoding parameters + strategies (part 3) * whisper : wip decoding parameters + strategies (part 4) * whisper : fix prompt_past update to not include prompt_init * whisper : temperature + best_of support * whisper : support for compression_ration_threshold We actually use entropy, but it is similar * command : fix example to use logits instead of obsolete probs * whisper : handle empty sequence ranking * whisper : add WHISPER_DEBUG + diagnostic prints + new main args * whisper : minor fixes * whisper : add beam-search support * whisper : bug fix when there no previous context * whisper : add comments * stream : disable temperature fallback For real-time processing, we always want a single decoder running at T=0 * whisper.swiftui : update example - fix paths + add empty folders	2023-01-15 11:29:57 +02:00
Georgi Gerganov	a6dbd9188b	stream : fix a bug that inserted a lot of empty audio at the start The quality was terrible due to this	2023-01-14 19:20:47 +02:00
Georgi Gerganov	4ef3398e8f	ggml : remove obsolete zeroing + comment fixes (#390 )	2023-01-08 20:21:03 +02:00
Ian Bicking	5e9f33596f	readme : clarify main and stream usage (#391 ) Give an example of ./main that uses a sample file that's already there, and make the stream example clarify you need `make stream`	2023-01-08 20:18:41 +02:00
Abitofevrything	8d7b29cedd	ggml : correct behaviour of ggml_vec_sum_f32 (#390 )	2023-01-08 20:06:09 +02:00
boolemancer	08dc705a69	whisper : fix sample_to_timestamp calculation with 64 bit precision to avoid overflow (#388 ) * Do calculation with 64 bit precision to avoid overflow * Update whisper.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-01-08 15:08:45 +02:00
Syahmi Azhar	1512545149	whisper : add loader class to allow loading from buffer and others (#353 ) * whisper : add loader to allow loading from other than file * whisper : rename whisper_init to whisper_init_from_file * whisper : add whisper_init_from_buffer * android : Delete local.properties * android : load models directly from assets * whisper : adding <stddef.h> needed for size_t + code style Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-01-08 13:03:33 +02:00
Georgi Gerganov	52a3e0c92a	ggml : improve vec_dot_f16 unrolling in flash_attn_f16	2023-01-08 11:41:18 +02:00
Georgi Gerganov	d1ea1220ff	command : clean-up / refactoring / formatting (#383 )	2023-01-07 21:43:24 +02:00
David	9c4a1522f6	command : always-prompt mode (#383 )	2023-01-07 21:41:11 +02:00
David Thorpe	f078a6f20e	go : adding features to the go-whisper example, go ci, etc (#384 ) * Updated bindings so they can be used in third pary packages. * Updated makefiles to set FMA flag on optionally, for xeon E5 on Darwin * Added test script * Changes for examples * Reverted * Made the NewContext method private	2023-01-07 21:21:43 +02:00
Georgi Gerganov	f30b5d322c	ggml : fix bug in new soft max computation	2023-01-07 21:00:07 +02:00
Georgi Gerganov	44efbf7ff1	cmake : add -Wno-unused-function + update whisper.js	2023-01-07 20:18:34 +02:00
Georgi Gerganov	d347a59a5f	ggml : when using BLAS start only 1 CPU thread	2023-01-07 19:48:56 +02:00
Georgi Gerganov	6394c906af	ggml : fix running tasks with variable number of threads	2023-01-07 19:20:18 +02:00
Georgi Gerganov	74ffa14e1d	ggml : unroll ggml_vec_dot_f16 in ggml_compute_forward_flash_attn_f16	2023-01-07 19:19:40 +02:00
Georgi Gerganov	65fdcbbbbb	whisper : revert accidental MB change	2023-01-07 16:18:21 +02:00
Georgi Gerganov	d61d55cd4b	ggml : speed-up soft max via Accelerate + unroll	2023-01-07 16:16:42 +02:00
Georgi Gerganov	d51fc3ee0a	ggml : use vDSP_sve and vDSP_maxv from Accelerate	2023-01-07 16:10:16 +02:00
Georgi Gerganov	f82a7dd019	ggml : make gcc happy (minor)	2023-01-07 09:34:39 +02:00
Georgi Gerganov	87dd4a3081	talk.wasm : bump memory usage + update whisper.js	2023-01-06 21:13:44 +02:00
m.bell	41e05c6b1b	cmake : support AVX2 in Windows better (#381 )	2023-01-06 19:36:33 +02:00
Georgi Gerganov	fa379cb22a	Revert "tmp" This reverts commit `1652965529`.	2023-01-06 19:33:09 +02:00
David Thorpe	322f4e6c4e	go : bindings updated so they can be used in third party packages. (#379 ) * Updated bindings so they can be used in third pary packages. * Updated makefiles to set FMA flag on optionally, for xeon E5 on Darwin	2023-01-06 19:32:28 +02:00
Georgi Gerganov	1652965529	tmp	2023-01-06 19:32:12 +02:00
Georgi Gerganov	6042c7a3be	cmake : change min required version to 3.0 (#351 ) We increase the min version only when want to use particular functionality that is available in the newer version	2023-01-06 19:25:28 +02:00
Georgi Gerganov	6b351bb669	command : add "guided-mode" video demo in the README.md	2023-01-06 18:59:26 +02:00
Abitofevrything	a62170c656	ggml : add SSE3 and fp16 conversion lookup table (#368 ) * Improves WASM performance: On MacBook M1 Pro, I observe 25% faster using Firefox and 35% faster using Chrome * Add support for SSE3 SIMD * Add SSE3 to system information * Add Imath support for fp16-fp32 conversions * Add Imath to system information * Wrap Imath calls to avoid static function warnings * Drop Imath; Add lookup table for f16 -> f32 conversions * Remove TODO comments * Update SSE3 to new macro arguments * Correct updated macro definitions * Prefer static inline where possible * ggml : static inlines + add public f16 <-> f32 conversions Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2023-01-06 18:45:59 +02:00
Thomas Fitzsimmons	1944e7c33e	whisper : document POWER VSX support	2023-01-05 23:53:00 +02:00
Thomas Fitzsimmons	49a8dd6732	ggml : reorganize POWER9 ppc64le SIMD code	2023-01-05 23:53:00 +02:00
Thomas Fitzsimmons	8c7f642286	ggml : change f16 load and store macro arguments	2023-01-05 23:53:00 +02:00
Georgi Gerganov	ad2a4ffa03	whisper : do not use F16 tensors when in F32 mode (#369 )	2023-01-05 22:56:25 +02:00
Georgi Gerganov	b3c865083e	ci : add emscripten build	2023-01-05 22:10:20 +02:00
Georgi Gerganov	a0d4f8e65c	main : make whisper_print_segment_callback() more readable (close #371 )	2023-01-05 21:45:05 +02:00
Georgi Gerganov	4a214d2f07	cmake : add CMAKE_RUNTIME_OUTPUT_DIRECTORY Currently needed by the wasm examples	2023-01-05 21:40:59 +02:00
Georgi Gerganov	0a0cfa7985	ggml : add void to argument-less functions	2023-01-05 21:40:38 +02:00
Georgi Gerganov	196d738974	minor : close #370 + Makefile build info print change	2023-01-05 21:35:45 +02:00
Andy Maloney	84c6b42e65	cmake : update to 3.19 (#351 ) - update from 3.0 (from 2014) to 3.19 (from 2020) - move some global setting onto the targets (through a cmake include)	2023-01-05 21:22:48 +02:00
Andy Maloney	dd6d582977	whisper : use ranged-based for loops for readability	2023-01-05 21:20:44 +02:00
Georgi Gerganov	d51c5eb906	ggml : define MIN / MAX only if not defined (minor)	2023-01-05 21:16:52 +02:00
Georgi Gerganov	0be6a1afd9	make : print build information	2023-01-02 13:35:26 +02:00
Georgi Gerganov	a466c3404d	stream : fix data race on bool + avoid division-by-zero	2023-01-02 10:20:50 +02:00
Georgi Gerganov	d629c034a4	models : fix HF model URL (close #356 )	2023-01-02 09:54:43 +02:00
Andy Maloney	f00509d57c	command : refactor to split command list & general transcription modes (#331 ) This makes it easier to understand if you're looking for only one of the capabilities.	2022-12-31 14:08:57 +02:00
Thomas Fitzsimmons	424c410c42	ggml : improve f16 acceleration for POWER9 ppc64le	2022-12-31 10:02:19 +02:00
Georgi Gerganov	d97e6005e9	whisper : add whisper_n_audio_ctx and check for invalid audio_ctx closes #344	2022-12-31 09:57:19 +02:00
Ikko Ashimine	3467230a77	models : fix typo in convert-h5-to-ggml.py signficant -> significant	2022-12-31 09:49:01 +02:00
Avik Sengupta	a091581eb3	cmake : add runtime destination install (#345 ) needed for mingw32 build to successfully install the dlls in the correct location	2022-12-31 09:48:00 +02:00
Georgi Gerganov	68daf6e487	whisper : avoid some memory allocations	2022-12-30 13:43:48 +02:00
Niels Mayer	a593b932e4	main : add -ocsv, aka --output-csv to output a CSV file Adds -ocsv, aka --output-csv feature to examples/main, which outputs a CSV file containing lines formatted as follows <startTime-in-integer-milliseconds>, <endTime-in-integer-milliseconds>, "<transcript-line-including-commas>".	2022-12-29 14:04:00 +02:00
Georgi Gerganov	9a8ad3db69	make : add i686 arch (close #329 )	2022-12-29 13:58:55 +02:00