Silero VAD Node

Minimal Node.js wrapper around the Silero VAD ONNX model, with a small CLI and parity tests against the Python implementation. The Node implementation runs VAD and silence stripping directly from ffmpeg streams to keep memory usage low on long files.

Install

npm install @jjhbw/silero-vad

Requires Node 18+ and ffmpeg available on PATH for decoding arbitrary audio formats.

Library usage

const {
  loadSileroVad,
  getSpeechTimestamps,
  writeStrippedAudio,
  WEIGHTS,
} = require("@jjhbw/silero-vad");

(async () => {
  const vad = await loadSileroVad("default"); // or WEIGHTS keys/custom path
  try {
    if (!vad.sampleRate) throw new Error("Model sample rate is undefined");
    const inputs = ["input.wav", "other.mp3"];
    for (const inputPath of inputs) {
      vad.resetStates(); // per file/stream
      const ts = await getSpeechTimestamps(inputPath, vad);
      // Each entry includes both seconds (start/end) and samples (startSample/endSample).
      console.log(inputPath, ts);
      // Example return value:
      // [
      //   { start: 0.36, end: 1.92, startSample: 5760, endSample: 30720 },
      //   { start: 2.41, end: 3.05, startSample: 38560, endSample: 48800 }
      // ]

      // Strip silences from the original file using the timestamps.
      // Pick any extension supported by ffmpeg (e.g., .wav, .flac).
      // Note: encoding speed varies by container/codec; uncompressed PCM (e.g., .wav) is fastest,
      // lossless compression (e.g., .flac) is slower, and lossy codecs (e.g., .mp3/.aac/.opus)
      // are typically the slowest to encode.
      const outPath = inputPath.replace(/\.[^.]+$/, ".stripped.wav");
      await writeStrippedAudio(inputPath, ts, vad.sampleRate, outPath);
    }
  } finally {
    await vad.session.release?.(); // once per process when shutting down
  }
})();

Guidelines:

• Load once, reuse: keep one SileroVad per concurrent worker.
• Call resetStates() before each new file/stream; the session and weights stay in memory.
• Call release() when shutting down.

CLI

npx @jjhbw/silero-vad --audio input.wav --audio other.mp3 [options]

Options:

• --model <key|path>: model key (default, 16k, 8k_16k, half, op18) or custom ONNX path (default: default, i.e., bundled 16k op15).
• --threshold <float>: speech probability threshold (default 0.5).
• --min-speech-ms <ms>: minimum speech duration in ms (default 250).
• --min-silence-ms <ms>: minimum silence duration in ms (default 100).
• --speech-pad-ms <ms>: padding added to each speech segment in ms (default 30).
• --time-resolution <n>: decimal places for seconds output (default 3).
• --neg-threshold <float>: override the negative threshold (default max(threshold - 0.15, 0.01)).
• --cps <float>: enable the timeline visualization and set chars per second (default 4).
• --strip-silence: write a new WAV file with silences removed.
• --output-dir <path>: output directory for strip-silence files (default: input dir).

Outputs an array of { file, timestamps } to stdout as JSON. The CLI reuses a single ONNX session and resets state per file. The sample rate is defined by the selected model (read from vad.sampleRate).

Development

Clone the repo to run benchmarks and tests locally.

Benchmark

git clone https://github.com/jjhbw/silero-vad
cd silero-vad/js-fork
npm install
node bench.js --audio data/test.mp3 --runs 5

The benchmark reports timings per file for streaming VAD and silence stripping. Stripped-audio files are written to a temporary directory and removed after each run.

Tests

Snapshot tests compare Node outputs against Python ground truth (tests/snapshots/onnx.json):

git clone https://github.com/jjhbw/silero-vad
cd silero-vad/js-fork
npm install
npm test

Ensure Python snapshots are generated (run pytest tests/test_snapshots.py in the repo root) and ffmpeg is installed.