Sherpa-ONNX Voice Activity Detection

Detect when someone is speaking versus silence in real time using decibri and sherpa-onnx Silero VAD. Runs entirely offline with no API key, no cloud service, and no network dependency.

What this does

This integration captures live microphone audio using decibri and feeds it to the Silero VAD model via sherpa-onnx. The VAD segments the audio stream into speech and non-speech regions, reporting when speech starts and ends with sample-level timestamps.

Choose this when you need to know when someone is speaking rather than what they are saying. Common uses include gating audio before sending to a speech recognizer, detecting turn-taking in conversations, or triggering recording only when speech is present.

Prerequisites

Install packages

Download the Silero VAD model

Download the Silero VAD ONNX model from the sherpa-onnx releases:

wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/silero_vad.onnx

This is a single small file (~2 MB).

Code walkthrough

1. Configuration

Configure the Silero VAD model. The key parameters are the detection threshold and the minimum silence/speech duration.

const Decibri = require('decibri');
const sherpa = require('sherpa-onnx');

const config = {
  sileroVad: {
    model: './silero_vad.onnx',
    threshold: 0.5,
    minSilenceDuration: 0.25,
    minSpeechDuration: 0.25,
    windowSize: 512,
  },
  sampleRate: 16000,
  debug: false,
  bufferSizeInSeconds: 60,
};

2. Create the VAD

Instantiate the voice activity detector.

const vad = new sherpa.Vad(config);

3. Open the microphone

Create a decibri instance at 16 kHz mono. The VAD processes audio in windows of 512 samples (32 ms at 16 kHz), but decibri delivers larger chunks which we feed in windows.

const mic = new Decibri({ sampleRate: 16000, channels: 1 });

4. Process audio and detect speech

Convert each incoming Int16 buffer to Float32, feed it to the VAD in window-sized chunks, and check for completed speech segments.

const windowSize = 512;
let speechActive = false;

mic.on('data', (chunk) => {
  // Convert Int16 PCM to Float32
  const int16 = new Int16Array(chunk.buffer, chunk.byteOffset, chunk.length / 2);
  const float32 = new Float32Array(int16.length);
  for (let i = 0; i < int16.length; i++) {
    float32[i] = int16[i] / 32768;
  }

  // Feed audio in window-sized chunks
  for (let offset = 0; offset + windowSize <= float32.length; offset += windowSize) {
    const window = float32.subarray(offset, offset + windowSize);
    vad.acceptWaveform(window);

    if (vad.isSpeechDetected() && !speechActive) {
      speechActive = true;
      console.log('[speech start]');
    }

    if (!vad.isSpeechDetected() && speechActive) {
      speechActive = false;
      console.log('[speech end]');
    }

    // Drain completed speech segments
    while (!vad.isEmpty()) {
      const segment = vad.front();
      const duration = (segment.samples.length / 16000).toFixed(2);
      console.log(`  segment: ${duration}s of speech`);
      vad.pop();
    }
  }
});

5. Clean shutdown

Stop the microphone when the user presses Ctrl+C.

process.on('SIGINT', () => {
  mic.stop();
  vad.free();
  process.exit(0);
});

console.log('Listening for speech... (Ctrl+C to stop)');

Full example

const Decibri = require('decibri');
const sherpa = require('sherpa-onnx');

const config = {
  sileroVad: {
    model: './silero_vad.onnx',
    threshold: 0.5,
    minSilenceDuration: 0.25,
    minSpeechDuration: 0.25,
    windowSize: 512,
  },
  sampleRate: 16000,
  debug: false,
  bufferSizeInSeconds: 60,
};

const vad = new sherpa.Vad(config);
const mic = new Decibri({ sampleRate: 16000, channels: 1 });

const windowSize = 512;
let speechActive = false;

mic.on('data', (chunk) => {
  const int16 = new Int16Array(chunk.buffer, chunk.byteOffset, chunk.length / 2);
  const float32 = new Float32Array(int16.length);
  for (let i = 0; i < int16.length; i++) {
    float32[i] = int16[i] / 32768;
  }

  for (let offset = 0; offset + windowSize <= float32.length; offset += windowSize) {
    const window = float32.subarray(offset, offset + windowSize);
    vad.acceptWaveform(window);

    if (vad.isSpeechDetected() && !speechActive) {
      speechActive = true;
      console.log('[speech start]');
    }

    if (!vad.isSpeechDetected() && speechActive) {
      speechActive = false;
      console.log('[speech end]');
    }

    while (!vad.isEmpty()) {
      const segment = vad.front();
      const duration = (segment.samples.length / 16000).toFixed(2);
      console.log(`  segment: ${duration}s of speech`);
      vad.pop();
    }
  }
});

process.on('SIGINT', () => {
  mic.stop();
  vad.free();
  process.exit(0);
});

console.log('Listening for speech... (Ctrl+C to stop)');