Voice activity detection

snakers4/silero-vad

Model ID: snakers4/silero-vad
Type: Voice activity detection

Use this model

Detect speech in a web app

Install the browser SDK:

npm install @wfloat/wfloat-web

Load the model and detect speech regions in an audio file:

import { loadVadModel } from "@wfloat/wfloat-web";

const vad = await loadVadModel("snakers4/silero-vad");

const file = document.querySelector<HTMLInputElement>("#audio-file")?.files?.[0];
if (!file) throw new Error("Choose an audio file first.");

const result = await vad.detect({ audio: file });

for (const segment of result.segments) {
  console.log(segment.startSec, segment.endSec);
}

For live detection, use vad.createSession({ onSpeechStart, onSpeechEnd }), then call session.startMicrophone() and session.stopMicrophone().

Detect speech in React Native

Install the React Native SDK:

npm install @wfloat/react-native-wfloat

Run pod install from the app’s ios directory, and configure microphone permission for the target platform. Then create a live VAD session:

import { loadVadModel } from '@wfloat/react-native-wfloat';

const vad = await loadVadModel('snakers4/silero-vad');
const session = await vad.createSession({
  onSpeechStart(event) {
    console.log('speech started near', event.startSec);
  },
  onSpeechEnd(segment) {
    console.log(segment.startSec, segment.endSec);
  },
});

await session.startMicrophone();

// Call these when detection is no longer needed.
await session.stopMicrophone();
await session.close();

For existing mono PCM samples, call vad.detect({ audio, sampleRate: 16000 }).

Detect speech in Python

Install the Python SDK:

pip install wfloat

Load the model and detect speech in a mono 16 kHz, 16-bit PCM WAV file:

import wfloat

vad = wfloat.load_vad_model(
    "snakers4/silero-vad",
    threshold=0.5,
    min_silence_duration_sec=0.5,
    min_speech_duration_sec=0.25,
)

result = vad.detect(audio="recording.wav")

for segment in result.segments:
    print(segment.start_sec, segment.end_sec)

For in-memory audio, pass a sequence of floating-point mono samples and set sample_rate=16000.