Realtime VAD

Realtime VAD is a realtime voice activity detection library for Node.js environments. It utilizes the Silero VAD and onnxruntime-node libraries to perform efficient speech detection.

Quickstart

Below is an example demonstrating how to use the library:

import { VoiceActivityDetector } from "realtime-vad";
import record from "node-record-lpcm16";

const SAMPLE_RATE = 16000;
const CHANNELS = 1;

async function main() {
  const vad = new VoiceActivityDetector({
    sampleRate: SAMPLE_RATE,
    channels: CHANNELS,
    frameDurationMs: 30,
    speechThreshold: 0.5,
    silenceDebounceMs: 600,
  });

  // Load the model
  await vad.init();

  // Listen to speech events
  vad.on("speechStart", (startTime) => {
    console.log("Speech start detected at", startTime);
  });

  vad.on("speechEnd", ({ start, end, duration }) => {
    console.log("Speech detected from", start, "to", end, "for", duration, "ms");
  });

  // Start recording from microphone
  const recording = record.record({
    sampleRate: SAMPLE_RATE,
    channels: CHANNELS,
    threshold: 0,
    audioType: "wav",
  });
  const micStream = recording.stream();

  let audioBuffer = Buffer.alloc(0);

  micStream.on("data", async (data) => {
    audioBuffer = Buffer.concat([audioBuffer, data]);
    while (audioBuffer.length >= vad.chunkBytes) {
      const frame = audioBuffer.subarray(0, vad.chunkBytes);
      audioBuffer = audioBuffer.subarray(vad.chunkBytes);
      await vad.processAudioChunk(frame);
    }
  });

  micStream.on("error", (err) => {
    console.error("Error with microphone stream:", err);
  });

  console.log("Listening for speech... Press CTRL+C to stop.");
}

main();

Installation

1. Clone the repository

   git clone https://github.com/ThEditor/realtime-vad.git
   cd realtime-vad

2. Install dependencies

   pnpm install

3. Build the project

   pnpm build

Background

This project started as an extension of ricky0123/vad-node to add real-time voice activity detection in Node.js (You can checkout the archived repository here).

As of October 2024, @ricky0123 has dropped support for ricky0123/vad-node in order to focus on ricky0123/vad-web. Following this change and inspired from @ricky0123, I've created this as a standalone voice activity detection library.

License

This project is licensed under the MIT License.

References

[1] Silero Team. (2021). Silero VAD: pre-trained enterprise-grade Voice Activity Detector (VAD), Number Detector and Language Classifier. GitHub, GitHub repository, https://github.com/snakers4/silero-vad, [email protected].