Add Golang API for VAD (#708)

.gitignore

@@ -86,3 +86,4 @@ vits-piper-*
 vits-coqui-*
 vits-mms-*
 *.tar.bz2
+sherpa-onnx-paraformer-trilingual-zh-cantonese-en

go-api-examples/README.md

@@ -6,4 +6,21 @@ Please refer to the documentation
 https://k2-fsa.github.io/sherpa/onnx/go-api/index.html
 for details.
 
+- [./non-streaming-decode-files](./non-streaming-decode-files) It shows how to use
+  a non-streaming ASR model to decode files
+
+- [./non-streaming-tts](./non-streaming-tts) It shows how to use a non-streaming TTS
+  model to convert text to speech
+
+- [./real-time-speech-recognition-from-microphone](./real-time-speech-recognition-from-microphone)
+  It shows how to use a streaming ASR model to recognize speech from a microphone in real-time
+
+- [./vad](./vad) It shows how to use silero VAD with Golang.
+
+- [./vad-asr-whisper](./vad-asr-whisper) It shows how to use silero VAD + Whisper
+  for speech recognition.
+
+- [./vad-asr-paraformer](./vad-asr-paraformer) It shows how to use silero VAD + Paraformer
+  for speech recognition.
+
 [sherpa-onnx]: https://github.com/k2-fsa/sherpa-onnx

go-api-examples/non-streaming-tts/main.go

@@ -57,8 +57,7 @@ func main() {
 	log.Println("Done!")
 
 	ok := audio.Save(filename)
-	if ok != 1 {
+	if !ok {
 		log.Fatalf("Failed to write", filename)
 	}
-
 }

go-api-examples/vad-asr-paraformer/go.mod

@@ -0,0 +1,3 @@
+module vad-asr-paraformer
+
+go 1.12

go-api-examples/vad-asr-paraformer/main.go

@@ -0,0 +1,147 @@
+package main
+
+import (
+	"fmt"
+	"github.com/gordonklaus/portaudio"
+	sherpa "github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx"
+	"log"
+	"strings"
+)
+
+func main() {
+	log.SetFlags(log.LstdFlags | log.Lmicroseconds)
+
+	// 1. Create VAD
+	config := sherpa.VadModelConfig{}
+
+	// Please download silero_vad.onnx from
+	// https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+
+	config.SileroVad.Model = "./silero_vad.onnx"
+	config.SileroVad.Threshold = 0.5
+	config.SileroVad.MinSilenceDuration = 0.5
+	config.SileroVad.MinSpeechDuration = 0.25
+	config.SileroVad.WindowSize = 512
+	config.SampleRate = 16000
+	config.NumThreads = 1
+	config.Provider = "cpu"
+	config.Debug = 1
+
+	var bufferSizeInSeconds float32 = 20
+
+	vad := sherpa.NewVoiceActivityDetector(&config, bufferSizeInSeconds)
+	defer sherpa.DeleteVoiceActivityDetector(vad)
+
+	// 2. Create ASR recognizer
+
+	c := sherpa.OfflineRecognizerConfig{}
+	c.FeatConfig.SampleRate = 16000
+	c.FeatConfig.FeatureDim = 80
+
+	// Please download the model from
+	// https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-paraformer-trilingual-zh-cantonese-en.tar.bz2
+	c.ModelConfig.Paraformer.Model = "./sherpa-onnx-paraformer-trilingual-zh-cantonese-en/model.int8.onnx"
+	c.ModelConfig.Tokens = "./sherpa-onnx-paraformer-trilingual-zh-cantonese-en/tokens.txt"
+	c.ModelConfig.NumThreads = 2
+	c.ModelConfig.Debug = 1
+	c.ModelConfig.Provider = "cpu"
+
+	recognizer := sherpa.NewOfflineRecognizer(&c)
+	defer sherpa.DeleteOfflineRecognizer(recognizer)
+
+	err := portaudio.Initialize()
+	if err != nil {
+		log.Fatalf("Unable to initialize portaudio: %v\n", err)
+	}
+	defer portaudio.Terminate()
+
+	default_device, err := portaudio.DefaultInputDevice()
+	if err != nil {
+		log.Fatal("Failed to get default input device: %v\n", err)
+	}
+	log.Printf("Selected default input device: %s\n", default_device.Name)
+	param := portaudio.StreamParameters{}
+	param.Input.Device = default_device
+	param.Input.Channels = 1
+	param.Input.Latency = default_device.DefaultHighInputLatency
+
+	param.SampleRate = float64(config.SampleRate)
+	param.FramesPerBuffer = 0
+	param.Flags = portaudio.ClipOff
+
+	// you can choose another value for 0.1 if you want
+	samplesPerCall := int32(param.SampleRate * 0.1) // 0.1 second
+	samples := make([]float32, samplesPerCall)
+
+	s, err := portaudio.OpenStream(param, samples)
+	if err != nil {
+		log.Fatalf("Failed to open the stream")
+	}
+
+	defer s.Close()
+	chk(s.Start())
+
+	log.Print("Started! Please speak")
+	printed := false
+
+	k := 0
+	for {
+		chk(s.Read())
+		vad.AcceptWaveform(samples)
+
+		if vad.IsSpeech() && !printed {
+			printed = true
+			log.Print("Detected speech\n")
+		}
+
+		if !vad.IsSpeech() {
+			printed = false
+		}
+
+		for !vad.IsEmpty() {
+			speechSegment := vad.Front()
+			vad.Pop()
+
+			duration := float32(len(speechSegment.Samples)) / float32(config.SampleRate)
+
+			audio := &sherpa.GeneratedAudio{}
+			audio.Samples = speechSegment.Samples
+			audio.SampleRate = config.SampleRate
+
+			// Now decode it
+			go decode(recognizer, audio, k)
+
+			k += 1
+
+			log.Printf("Duration: %.2f seconds\n", duration)
+		}
+	}
+
+	chk(s.Stop())
+}
+
+func decode(recognizer *sherpa.OfflineRecognizer, audio *sherpa.GeneratedAudio, id int) {
+	stream := sherpa.NewOfflineStream(recognizer)
+	defer sherpa.DeleteOfflineStream(stream)
+	stream.AcceptWaveform(audio.SampleRate, audio.Samples)
+	recognizer.Decode(stream)
+	result := stream.GetResult()
+	text := strings.ToLower(result.Text)
+	text = strings.Trim(text, " ")
+	log.Println(text)
+
+	duration := float32(len(audio.Samples)) / float32(audio.SampleRate)
+
+	filename := fmt.Sprintf("seg-%d-%.2f-seconds-%s.wav", id, duration, text)
+	ok := audio.Save(filename)
+	if ok {
+		log.Printf("Saved to %s", filename)
+	}
+	log.Print("----------\n")
+}
+
+func chk(err error) {
+	if err != nil {
+		panic(err)
+	}
+}

go-api-examples/vad-asr-paraformer/run.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+
+if [ ! -f ./silero_vad.onnx ]; then
+  curl -SL -O https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+fi
+
+if [ ! -f ./sherpa-onnx-paraformer-trilingual-zh-cantonese-en/model.int8.onnx ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-paraformer-trilingual-zh-cantonese-en.tar.bz2
+  tar xvf sherpa-onnx-paraformer-trilingual-zh-cantonese-en.tar.bz2
+  rm sherpa-onnx-paraformer-trilingual-zh-cantonese-en.tar.bz2
+fi
+
+go mod tidy
+go build
+./vad-asr-paraformer

go-api-examples/vad-asr-whisper/go.mod

@@ -0,0 +1,3 @@
+module vad-asr-whisper
+
+go 1.12

go-api-examples/vad-asr-whisper/main.go

@@ -0,0 +1,145 @@
+package main
+
+import (
+	"fmt"
+	"github.com/gordonklaus/portaudio"
+	sherpa "github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx"
+	"log"
+	"strings"
+)
+
+func main() {
+	log.SetFlags(log.LstdFlags | log.Lmicroseconds)
+
+	// 1. Create VAD
+	config := sherpa.VadModelConfig{}
+
+	// Please download silero_vad.onnx from
+	// https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+
+	config.SileroVad.Model = "./silero_vad.onnx"
+	config.SileroVad.Threshold = 0.5
+	config.SileroVad.MinSilenceDuration = 0.5
+	config.SileroVad.MinSpeechDuration = 0.25
+	config.SileroVad.WindowSize = 512
+	config.SampleRate = 16000
+	config.NumThreads = 1
+	config.Provider = "cpu"
+	config.Debug = 1
+
+	var bufferSizeInSeconds float32 = 20
+
+	vad := sherpa.NewVoiceActivityDetector(&config, bufferSizeInSeconds)
+	defer sherpa.DeleteVoiceActivityDetector(vad)
+
+	// 2. Create ASR recognizer
+
+	c := sherpa.OfflineRecognizerConfig{}
+	c.FeatConfig.SampleRate = 16000
+	c.FeatConfig.FeatureDim = 80
+	c.ModelConfig.Whisper.Encoder = "./sherpa-onnx-whisper-tiny.en/tiny.en-encoder.int8.onnx"
+	c.ModelConfig.Whisper.Decoder = "./sherpa-onnx-whisper-tiny.en/tiny.en-decoder.int8.onnx"
+	c.ModelConfig.Tokens = "./sherpa-onnx-whisper-tiny.en/tiny.en-tokens.txt"
+	c.ModelConfig.NumThreads = 2
+	c.ModelConfig.Debug = 1
+	c.ModelConfig.Provider = "cpu"
+
+	recognizer := sherpa.NewOfflineRecognizer(&c)
+	defer sherpa.DeleteOfflineRecognizer(recognizer)
+
+	err := portaudio.Initialize()
+	if err != nil {
+		log.Fatalf("Unable to initialize portaudio: %v\n", err)
+	}
+	defer portaudio.Terminate()
+
+	default_device, err := portaudio.DefaultInputDevice()
+	if err != nil {
+		log.Fatal("Failed to get default input device: %v\n", err)
+	}
+	log.Printf("Selected default input device: %s\n", default_device.Name)
+	param := portaudio.StreamParameters{}
+	param.Input.Device = default_device
+	param.Input.Channels = 1
+	param.Input.Latency = default_device.DefaultHighInputLatency
+
+	param.SampleRate = float64(config.SampleRate)
+	param.FramesPerBuffer = 0
+	param.Flags = portaudio.ClipOff
+
+	// you can choose another value for 0.1 if you want
+	samplesPerCall := int32(param.SampleRate * 0.1) // 0.1 second
+	samples := make([]float32, samplesPerCall)
+
+	s, err := portaudio.OpenStream(param, samples)
+	if err != nil {
+		log.Fatalf("Failed to open the stream")
+	}
+
+	defer s.Close()
+	chk(s.Start())
+
+	log.Print("Started! Please speak")
+	printed := false
+
+	k := 0
+	for {
+		chk(s.Read())
+		vad.AcceptWaveform(samples)
+
+		if vad.IsSpeech() && !printed {
+			printed = true
+			log.Print("Detected speech\n")
+		}
+
+		if !vad.IsSpeech() {
+			printed = false
+		}
+
+		for !vad.IsEmpty() {
+			speechSegment := vad.Front()
+			vad.Pop()
+
+			duration := float32(len(speechSegment.Samples)) / float32(config.SampleRate)
+
+			audio := &sherpa.GeneratedAudio{}
+			audio.Samples = speechSegment.Samples
+			audio.SampleRate = config.SampleRate
+
+			// Now decode it
+			go decode(recognizer, audio, k)
+
+			k += 1
+
+			log.Printf("Duration: %.2f seconds\n", duration)
+		}
+	}
+
+	chk(s.Stop())
+}
+
+func decode(recognizer *sherpa.OfflineRecognizer, audio *sherpa.GeneratedAudio, id int) {
+	stream := sherpa.NewOfflineStream(recognizer)
+	defer sherpa.DeleteOfflineStream(stream)
+	stream.AcceptWaveform(audio.SampleRate, audio.Samples)
+	recognizer.Decode(stream)
+	result := stream.GetResult()
+	text := strings.ToLower(result.Text)
+	text = strings.Trim(text, " ")
+	log.Println(text)
+
+	duration := float32(len(audio.Samples)) / float32(audio.SampleRate)
+
+	filename := fmt.Sprintf("seg-%d-%.2f-seconds-%s.wav", id, duration, text)
+	ok := audio.Save(filename)
+	if ok {
+		log.Printf("Saved to %s", filename)
+	}
+	log.Print("----------\n")
+}
+
+func chk(err error) {
+	if err != nil {
+		panic(err)
+	}
+}

go-api-examples/vad-asr-whisper/run.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+
+
+if [ ! -f ./silero_vad.onnx ]; then
+  curl -SL -O https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+fi
+
+if [ ! -f ./sherpa-onnx-whisper-tiny.en/tiny.en-encoder.int8.onnx ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-tiny.en.tar.bz2
+  tar xvf sherpa-onnx-whisper-tiny.en.tar.bz2
+  rm sherpa-onnx-whisper-tiny.en.tar.bz2
+fi
+
+go mod tidy
+go build
+./vad-asr-whisper

go-api-examples/vad/go.mod

@@ -0,0 +1,3 @@
+module vad
+
+go 1.12

go-api-examples/vad/main.go

@@ -0,0 +1,112 @@
+package main
+
+import (
+	"fmt"
+	"github.com/gordonklaus/portaudio"
+	sherpa "github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx"
+	"log"
+)
+
+func main() {
+	log.SetFlags(log.LstdFlags | log.Lmicroseconds)
+
+	config := sherpa.VadModelConfig{}
+
+	// Please download silero_vad.onnx from
+	// https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+
+	config.SileroVad.Model = "./silero_vad.onnx"
+	config.SileroVad.Threshold = 0.5
+	config.SileroVad.MinSilenceDuration = 0.5
+	config.SileroVad.MinSpeechDuration = 0.25
+	config.SileroVad.WindowSize = 512
+	config.SampleRate = 16000
+	config.NumThreads = 1
+	config.Provider = "cpu"
+	config.Debug = 1
+
+	var bufferSizeInSeconds float32 = 5
+
+	vad := sherpa.NewVoiceActivityDetector(&config, bufferSizeInSeconds)
+	defer sherpa.DeleteVoiceActivityDetector(vad)
+
+	err := portaudio.Initialize()
+	if err != nil {
+		log.Fatalf("Unable to initialize portaudio: %v\n", err)
+	}
+	defer portaudio.Terminate()
+
+	default_device, err := portaudio.DefaultInputDevice()
+	if err != nil {
+		log.Fatal("Failed to get default input device: %v\n", err)
+	}
+	log.Printf("Selected default input device: %s\n", default_device.Name)
+	param := portaudio.StreamParameters{}
+	param.Input.Device = default_device
+	param.Input.Channels = 1
+	param.Input.Latency = default_device.DefaultLowInputLatency
+
+	param.SampleRate = float64(config.SampleRate)
+	param.FramesPerBuffer = 0
+	param.Flags = portaudio.ClipOff
+
+	// you can choose another value for 0.1 if you want
+	samplesPerCall := int32(param.SampleRate * 0.1) // 0.1 second
+	samples := make([]float32, samplesPerCall)
+
+	s, err := portaudio.OpenStream(param, samples)
+	if err != nil {
+		log.Fatalf("Failed to open the stream")
+	}
+
+	defer s.Close()
+	chk(s.Start())
+
+	log.Print("Started! Please speak")
+	printed := false
+
+	k := 0
+	for {
+		chk(s.Read())
+		vad.AcceptWaveform(samples)
+
+		if vad.IsSpeech() && !printed {
+			printed = true
+			log.Print("Detected speech\n")
+		}
+
+		if !vad.IsSpeech() {
+			printed = false
+		}
+
+		for !vad.IsEmpty() {
+			speechSegment := vad.Front()
+			vad.Pop()
+
+			duration := float32(len(speechSegment.Samples)) / float32(config.SampleRate)
+
+			audio := sherpa.GeneratedAudio{}
+			audio.Samples = speechSegment.Samples
+			audio.SampleRate = config.SampleRate
+
+			filename := fmt.Sprintf("seg-%d-%.2f-seconds.wav", k, duration)
+			ok := audio.Save(filename)
+			if ok {
+				log.Printf("Saved to %s", filename)
+			}
+
+			k += 1
+
+			log.Printf("Duration: %.2f seconds\n", duration)
+			log.Print("----------\n")
+		}
+	}
+
+	chk(s.Stop())
+}
+
+func chk(err error) {
+	if err != nil {
+		panic(err)
+	}
+}

go-api-examples/vad/run.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+
+if [ ! -f ./silero_vad.onnx ]; then
+  curl -SL -O https://github.com/snakers4/silero-vad/blob/master/files/silero_vad.onnx
+fi
+
+go mod tidy
+go build
+./vad

scripts/apk/generate-tts-apk-script.py

@@ -234,6 +234,12 @@ def get_mimic3_models() -> List[TtsModel]:
 def get_vits_models() -> List[TtsModel]:
     return [
         # Chinese
+        TtsModel(
+            model_dir="vits-icefall-zh-aishell3",
+            model_name="model.onnx",
+            lang="zh",
+            rule_fsts="vits-icefall-zh-aishell3/phone.fst,vits-icefall-zh-aishell3/date.fst,vits-icefall-zh-aishell3/rule.fst",
+        ),
         TtsModel(
             model_dir="vits-zh-aishell3",
             model_name="vits-aishell3.onnx",

scripts/go/_internal/.gitignore

@@ -1 +1,2 @@
 !*.sh
+go.sum

scripts/go/_internal/vad-asr-paraformer/.gitignore

@@ -0,0 +1 @@
+vad-asr-paraformer

scripts/go/_internal/vad-asr-paraformer/go.mod

@@ -0,0 +1,10 @@
+module vad-asr-paraformer
+
+go 1.12
+
+replace github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx => ../
+
+require (
+	github.com/gordonklaus/portaudio v0.0.0-20230709114228-aafa478834f5
+	github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx v0.0.0-00010101000000-000000000000
+)

scripts/go/_internal/vad-asr-paraformer/main.go

@@ -0,0 +1 @@
+../../../../go-api-examples/vad-asr-paraformer/main.go

scripts/go/_internal/vad-asr-paraformer/run.sh

@@ -0,0 +1 @@
+../../../../go-api-examples/vad-asr-paraformer/run.sh

scripts/go/_internal/vad-asr-whisper/.gitignore

@@ -0,0 +1,2 @@
+vad-asr-whisper
+

scripts/go/_internal/vad-asr-whisper/go.mod

@@ -0,0 +1,5 @@
+module vad-asr-whisper
+
+go 1.12
+
+replace github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx => ../

scripts/go/_internal/vad-asr-whisper/main.go

@@ -0,0 +1 @@
+../../../../go-api-examples/vad-asr-whisper/main.go

scripts/go/_internal/vad-asr-whisper/run.sh

@@ -0,0 +1 @@
+../../../../go-api-examples/vad-asr-whisper/run.sh

scripts/go/_internal/vad/.gitignore

@@ -0,0 +1,2 @@
+vad
+go.sum

scripts/go/_internal/vad/go.mod

@@ -0,0 +1,5 @@
+module vad
+
+go 1.12
+
+replace github.com/k2-fsa/sherpa-onnx-go/sherpa_onnx => ../

scripts/go/_internal/vad/main.go

@@ -0,0 +1 @@
+../../../../go-api-examples/vad/main.go

scripts/go/_internal/vad/run.sh

@@ -0,0 +1 @@
+../../../../go-api-examples/vad/run.sh

scripts/go/sherpa_onnx.go

@@ -614,6 +614,9 @@ func (tts *OfflineTts) Generate(text string, sid int, speed float32) *GeneratedA
 	ans.SampleRate = int(audio.sample_rate)
 	n := int(audio.n)
 	ans.Samples = make([]float32, n)
+
+	// see https://stackoverflow.com/questions/48756732/what-does-1-30c-yourtype-do-exactly-in-cgo
+	// :n:n means 0:n:n, means low:high:capacity
 	samples := (*[1 << 28]C.float)(unsafe.Pointer(audio.samples))[:n:n]
 	// copy(ans.Samples, samples)
 	for i := 0; i < n; i++ {
@@ -623,11 +626,160 @@ func (tts *OfflineTts) Generate(text string, sid int, speed float32) *GeneratedA
 	return ans
 }
 
-func (audio *GeneratedAudio) Save(filename string) int {
+func (audio *GeneratedAudio) Save(filename string) bool {
 	s := C.CString(filename)
 	defer C.free(unsafe.Pointer(s))
 
 	ok := int(C.SherpaOnnxWriteWave((*C.float)(&audio.Samples[0]), C.int(len(audio.Samples)), C.int(audio.SampleRate), s))
 
-	return ok
+	return ok == 1
+}
+
+// ============================================================
+// For VAD
+// ============================================================
+type SileroVadModelConfig struct {
+	Model              string
+	Threshold          float32
+	MinSilenceDuration float32
+	MinSpeechDuration  float32
+	WindowSize         int
+}
+
+type VadModelConfig struct {
+	SileroVad  SileroVadModelConfig
+	SampleRate int
+	NumThreads int
+	Provider   string
+	Debug      int
+}
+
+type CircularBuffer struct {
+	impl *C.struct_SherpaOnnxCircularBuffer
+}
+
+func DeleteCircularBuffer(buffer *CircularBuffer) {
+	C.SherpaOnnxDestroyCircularBuffer(buffer.impl)
+	buffer.impl = nil
+}
+
+func NewCircularBuffer(capacity int) *CircularBuffer {
+	circularBuffer := &CircularBuffer{}
+	circularBuffer.impl = C.SherpaOnnxCreateCircularBuffer(C.int(capacity))
+	return circularBuffer
+}
+
+func (buffer *CircularBuffer) Push(samples []float32) {
+	C.SherpaOnnxCircularBufferPush(buffer.impl, (*C.float)(&samples[0]), C.int(len(samples)))
+}
+
+func (buffer *CircularBuffer) Get(start int, n int) []float32 {
+	samples := C.SherpaOnnxCircularBufferGet(buffer.impl, C.int(start), C.int(n))
+	defer C.SherpaOnnxCircularBufferFree(samples)
+
+	result := make([]float32, n)
+
+	p := (*[1 << 28]C.float)(unsafe.Pointer(samples))[:n:n]
+	for i := 0; i < n; i++ {
+		result[i] = float32(p[i])
+	}
+
+	return result
+}
+
+func (buffer *CircularBuffer) Pop(n int) {
+	C.SherpaOnnxCircularBufferPop(buffer.impl, C.int(n))
+}
+
+func (buffer *CircularBuffer) Size() int {
+	return int(C.SherpaOnnxCircularBufferSize(buffer.impl))
+}
+
+func (buffer *CircularBuffer) Head() int {
+	return int(C.SherpaOnnxCircularBufferHead(buffer.impl))
+}
+
+func (buffer *CircularBuffer) Reset() {
+	C.SherpaOnnxCircularBufferReset(buffer.impl)
+}
+
+type SpeechSegment struct {
+	Start   int
+	Samples []float32
+}
+
+type VoiceActivityDetector struct {
+	impl *C.struct_SherpaOnnxVoiceActivityDetector
+}
+
+func NewVoiceActivityDetector(config *VadModelConfig, bufferSizeInSeconds float32) *VoiceActivityDetector {
+	c := C.struct_SherpaOnnxVadModelConfig{}
+
+	c.silero_vad.model = C.CString(config.SileroVad.Model)
+	defer C.free(unsafe.Pointer(c.silero_vad.model))
+
+	c.silero_vad.threshold = C.float(config.SileroVad.Threshold)
+	c.silero_vad.min_silence_duration = C.float(config.SileroVad.MinSilenceDuration)
+	c.silero_vad.min_speech_duration = C.float(config.SileroVad.MinSpeechDuration)
+	c.silero_vad.window_size = C.int(config.SileroVad.WindowSize)
+
+	c.sample_rate = C.int(config.SampleRate)
+	c.num_threads = C.int(config.NumThreads)
+	c.provider = C.CString(config.Provider)
+	defer C.free(unsafe.Pointer(c.provider))
+
+	c.debug = C.int(config.Debug)
+
+	vad := &VoiceActivityDetector{}
+	vad.impl = C.SherpaOnnxCreateVoiceActivityDetector(&c, C.float(bufferSizeInSeconds))
+
+	return vad
+}
+
+func DeleteVoiceActivityDetector(vad *VoiceActivityDetector) {
+	C.SherpaOnnxDestroyVoiceActivityDetector(vad.impl)
+	vad.impl = nil
+}
+
+func (vad *VoiceActivityDetector) AcceptWaveform(samples []float32) {
+	C.SherpaOnnxVoiceActivityDetectorAcceptWaveform(vad.impl, (*C.float)(&samples[0]), C.int(len(samples)))
+}
+
+func (vad *VoiceActivityDetector) IsEmpty() bool {
+	return 1 == int(C.SherpaOnnxVoiceActivityDetectorEmpty(vad.impl))
+}
+
+func (vad *VoiceActivityDetector) IsSpeech() bool {
+	return 1 == int(C.SherpaOnnxVoiceActivityDetectorDetected(vad.impl))
+}
+
+func (vad *VoiceActivityDetector) Pop() {
+	C.SherpaOnnxVoiceActivityDetectorPop(vad.impl)
+}
+
+func (vad *VoiceActivityDetector) Clear() {
+	C.SherpaOnnxVoiceActivityDetectorClear(vad.impl)
+}
+
+func (vad *VoiceActivityDetector) Front() *SpeechSegment {
+	f := C.SherpaOnnxVoiceActivityDetectorFront(vad.impl)
+	defer C.SherpaOnnxDestroySpeechSegment(f)
+
+	ans := &SpeechSegment{}
+	ans.Start = int(f.start)
+
+	n := int(f.n)
+	ans.Samples = make([]float32, n)
+
+	samples := (*[1 << 28]C.float)(unsafe.Pointer(f.samples))[:n:n]
+
+	for i := 0; i < n; i++ {
+		ans.Samples[i] = float32(samples[i])
+	}
+
+	return ans
+}
+
+func (vad *VoiceActivityDetector) Reset() {
+	C.SherpaOnnxVoiceActivityDetectorReset(vad.impl)
 }

sherpa-onnx/c-api/c-api.cc

@@ -309,6 +309,9 @@ SherpaOnnxOfflineRecognizer *CreateOfflineRecognizer(
 
   recognizer_config.model_config.whisper.task =
       SHERPA_ONNX_OR(config->model_config.whisper.task, "transcribe");
+  if (recognizer_config.model_config.whisper.task.empty()) {
+    recognizer_config.model_config.whisper.task = "transcribe";
+  }
 
   recognizer_config.model_config.tdnn.model =
       SHERPA_ONNX_OR(config->model_config.tdnn.model, "");
@@ -331,6 +334,11 @@ SherpaOnnxOfflineRecognizer *CreateOfflineRecognizer(
 
   recognizer_config.decoding_method =
       SHERPA_ONNX_OR(config->decoding_method, "greedy_search");
+
+  if (recognizer_config.decoding_method.empty()) {
+    recognizer_config.decoding_method = "greedy_search";
+  }
+
   recognizer_config.max_active_paths =
       SHERPA_ONNX_OR(config->max_active_paths, 4);