auditok-webrtcvad

The WebRTC voice activity detector as a frame validator for auditok: detect speech specifically, rather than any acoustic activity, and let auditok's tokenizer turn the frame decisions into events with proper duration and silence semantics — the JS counterpart of Python auditok's validator="webrtc:N".

This is libfvad — the standalone, BSD-licensed extraction of WebRTC's VAD, a small GMM over frequency sub-band energies with online noise adaptation — compiled once to a ~30 KB WebAssembly binary embedded in the package. No model downloads, no native modules, no dependencies; the same file runs in Node and every browser. It wraps the identical C code as Python's webrtcvad package, and the test suite verifies frame decisions match it bit for bit.

Install

npm install auditok auditok-webrtcvad

The main auditok package stays a few KB of plain JS; this add-on is the explicit opt-in for the WASM blob.

Usage

import { split } from "auditok";
import { createWebrtcValidator } from "auditok-webrtcvad";

const validator = await createWebrtcValidator({ sampleRate: 16000, mode: 1 });
for await (const event of split(audio, { validator, maxSilence: 0.1 })) {
  console.log(`speech: ${event.start.toFixed(2)}s -> ${event.end.toFixed(2)}s`);
}
validator.destroy(); // optional; also released on garbage collection

Options of createWebrtcValidator:

• sampleRate (required): 8000, 16000, 32000 or 48000 Hz — a WebRTC VAD requirement; must match the audio given to split.
• mode (default 1): aggressiveness 0–3. Higher rejects more audio as non-speech; 0–1 suit far-field or noisy audio, 2 clean close-talk audio.
• subframeDur (default 0.01): the VAD consumes 10, 20 or 30 ms subframes; the analysisWindow used with split should be a multiple.
• aggregation (default "majority"): how subframe decisions combine into a window decision ("majority", "any", "all").
• useChannel (default "mix"): channel index or "mix" for multichannel audio.

The VAD is stateful (it adapts its noise model and applies a decision hangover): use a fresh validator per stream or file, or call .reset() between unrelated inputs. As a frame validator it typically works best with a smaller maxSilence than auditok's 0.3 s default — try 0.1.

For raw per-subframe decisions (the equivalent of Python's webrtcvad.Vad), use the lower-level WebrtcVad class.

Rebuilding the WASM binary

src/fvad-wasm.ts is generated and committed; users and CI never need a C toolchain. To rebuild it (e.g. to bump the pinned libfvad commit), run tools/build-fvad-wasm.sh from the repository root (requires docker).

License

MIT (wrapper) and BSD-3-Clause (libfvad / WebRTC — see LICENSE.libfvad).