Add VAD + Non-streaming ASR example for JavaScript API. (#1170)

nodejs-addon-examples/README.md

@@ -93,6 +93,7 @@ The following tables list the examples in this folder.
 |---|---|
 |[./test_asr_non_streaming_transducer.js](./test_asr_non_streaming_transducer.js)|Non-streaming speech recognition from a file with a Zipformer transducer model|
 |[./test_asr_non_streaming_whisper.js](./test_asr_non_streaming_whisper.js)| Non-streaming speech recognition from a file using [Whisper](https://github.com/openai/whisper)|
+|[./test_vad_with_non_streaming_asr_whisper.js](./test_vad_with_non_streaming_asr_whisper.js)| Non-streaming speech recognition from a file using [Whisper](https://github.com/openai/whisper) + [Silero VAD](https://github.com/snakers4/silero-vad)|
 |[./test_asr_non_streaming_nemo_ctc.js](./test_asr_non_streaming_nemo_ctc.js)|Non-streaming speech recognition from a file using a [NeMo](https://github.com/NVIDIA/NeMo) CTC model with greedy search|
 |[./test_asr_non_streaming_paraformer.js](./test_asr_non_streaming_paraformer.js)|Non-streaming speech recognition from a file using [Paraformer](https://github.com/alibaba-damo-academy/FunASR)|
 |[./test_asr_non_streaming_sense_voice.js](./test_asr_non_streaming_sense_voice.js)|Non-streaming speech recognition from a file using [SenseVoice](https://github.com/FunAudioLLM/SenseVoice)|
@@ -221,11 +222,24 @@ rm sherpa-onnx-whisper-tiny.en.tar.bz2
 
 node ./test_asr_non_streaming_whisper.js
 
-# To run VAD + non-streaming ASR with Paraformer using a microphone
+# To run VAD + non-streaming ASR with Whisper using a microphone
 npm install naudiodon2
 node ./test_vad_asr_non_streaming_whisper_microphone.js
 ```
 
+### Non-streaming speech recognition with Whisper + VAD
+
+```bash
+wget 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
+
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/Obama.wav
+wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx
+
+node ./test_vad_with_non_streaming_asr_whisper.js
+```
+
 ### Non-streaming speech recognition with NeMo CTC models
 
 ```bash

nodejs-addon-examples/package.json

@@ -1,5 +1,5 @@
 {
   "dependencies": {
-    "sherpa-onnx-node": "^1.10.17"
+    "sherpa-onnx-node": "^1.10.18"
   }
 }

nodejs-addon-examples/test_vad_with_non_streaming_asr_whisper.js

@@ -0,0 +1,127 @@
+// Copyright (c)  2023-2024  Xiaomi Corporation (authors: Fangjun Kuang)
+
+const sherpa_onnx = require('sherpa-onnx-node');
+
+function createRecognizer() {
+  // Please download test files from
+  // https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+  const config = {
+    'featConfig': {
+      'sampleRate': 16000,
+      'featureDim': 80,
+    },
+    'modelConfig': {
+      'whisper': {
+        'encoder': './sherpa-onnx-whisper-tiny.en/tiny.en-encoder.int8.onnx',
+        'decoder': './sherpa-onnx-whisper-tiny.en/tiny.en-decoder.int8.onnx',
+      },
+      'tokens': './sherpa-onnx-whisper-tiny.en/tiny.en-tokens.txt',
+      'numThreads': 2,
+      'provider': 'cpu',
+      'debug': 1,
+    }
+  };
+
+  return new sherpa_onnx.OfflineRecognizer(config);
+}
+
+function createVad() {
+  // please download silero_vad.onnx from
+  // https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx
+  const config = {
+    sileroVad: {
+      model: './silero_vad.onnx',
+      threshold: 0.5,
+      minSpeechDuration: 0.25,
+      minSilenceDuration: 0.5,
+      windowSize: 512,
+    },
+    sampleRate: 16000,
+    debug: true,
+    numThreads: 1,
+  };
+
+  const bufferSizeInSeconds = 60;
+
+  return new sherpa_onnx.Vad(config, bufferSizeInSeconds);
+}
+
+const recognizer = createRecognizer();
+const vad = createVad();
+
+// please download ./Obama.wav from
+// https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
+const waveFilename = './Obama.wav';
+const wave = sherpa_onnx.readWave(waveFilename);
+
+if (wave.sampleRate != recognizer.config.featConfig.sampleRate) {
+  throw new Error(
+      'Expected sample rate: ${recognizer.config.featConfig.sampleRate}. Given: ${wave.sampleRate}');
+}
+
+console.log('Started')
+let start = Date.now();
+
+const windowSize = vad.config.sileroVad.windowSize;
+for (let i = 0; i < wave.samples.length; i += windowSize) {
+  const thisWindow = wave.samples.subarray(i, i + windowSize);
+  vad.acceptWaveform(thisWindow);
+
+  while (!vad.isEmpty()) {
+    const segment = vad.front();
+    vad.pop();
+
+    let start_time = segment.start / wave.sampleRate;
+    let end_time = start_time + segment.samples.length / wave.sampleRate;
+
+    start_time = start_time.toFixed(2);
+    end_time = end_time.toFixed(2);
+
+    const stream = recognizer.createStream();
+    stream.acceptWaveform(
+        {samples: segment.samples, sampleRate: wave.sampleRate});
+
+    recognizer.decode(stream);
+    const r = recognizer.getResult(stream);
+    if (r.text.length > 0) {
+      const text = r.text.toLowerCase().trim();
+      console.log(`${start_time} -- ${end_time}: ${text}`);
+    }
+  }
+}
+
+vad.flush();
+
+while (!vad.isEmpty()) {
+  const segment = vad.front();
+  vad.pop();
+
+  let start_time = segment.start / wave.sampleRate;
+  let end_time = start_time + segment.samples.length / wave.sampleRate;
+
+  start_time = start_time.toFixed(2);
+  end_time = end_time.toFixed(2);
+
+  const stream = recognizer.createStream();
+  stream.acceptWaveform(
+      {samples: segment.samples, sampleRate: wave.sampleRate});
+
+  recognizer.decode(stream);
+  const r = recognizer.getResult(stream);
+  if (r.text.length > 0) {
+    const text = r.text.toLowerCase().trim();
+    console.log(`${start_time} -- ${end_time}: ${text}`);
+  }
+}
+
+let stop = Date.now();
+console.log('Done')
+
+const elapsed_seconds = (stop - start) / 1000;
+const duration = wave.samples.length / wave.sampleRate;
+const real_time_factor = elapsed_seconds / duration;
+console.log('Wave duration', duration.toFixed(3), 'secodns')
+console.log('Elapsed', elapsed_seconds.toFixed(3), 'secodns')
+console.log(
+    `RTF = ${elapsed_seconds.toFixed(3)}/${duration.toFixed(3)} =`,
+    real_time_factor.toFixed(3))