webrtcvad is a cross-platform, native node.js addon for detecting speech in raw audio. webrtcvad provides node.js bindings for the native WebRTC voice activity detection library.
webrtcvad has been tested on Windows 10, macOS 10.14+, and Ubuntu 18.04+ (and may work on other platforms as well).
To install webrtcvad, run:
yarn add webrtcvad
If you're using this library with Electron, you should probably use electron-rebuild.
webrtcvad works with single-channel, 16-bit, signed-integer linear PCM audio sampled at 8kHz, 16kHz, 32kHz, or 48kHz.
You can create a voice activity detector (VAD) with:
import VAD from "webrtcvad";
const audio = ...; // Buffer containing exactly 10ms, 20ms, or 30ms of audio data
const vad = new VAD(16000, 3);
vad.process(audio);
The VAD constructor accepts two arguments. The first is the sample rate (which must be one of the above), and the second is a VAD level, with 0 being the least aggressive (i.e., leading to the most false positives) and 3 being the most aggressive (i.e., leading to the least false positives, but possibly missing actual speech).
The VAD object has one method called process that returns true if the frame contains speech and false if not. It takes one argument, a Buffer containing exactly 10ms, 20ms, or 30ms of audio. For instance, if you're sampling at 16kHz, then the length of audio should be 160, 320, or 480, since the audio must be single-channel, 16-bit.
If you're looking to apply the VAD to live microphone data, take a look at speech-recorder, which uses webrtcvad to capture speech from a device's microphone.
This module is based on py-webrtcvad.