← All models

voice-activity-detection · audio · WASM (onnxruntime-web) · ~2 MB

Silero VAD — find the speech in any audio

A voice activity detector answers one small, useful question, thirty-odd times a second: is someone speaking right now? Silero VAD is a tiny (~2 MB) recurrent network that reads audio in 512-sample frames and emits a speech probability for each one. It's the quiet workhorse in front of almost every speech system — it decides what to send to a transcriber, when to stop recording, and where the pauses are. Here it runs entirely on your device.

At a glance

Modelonnx-community/silero-vad (Silero VAD v5)
Taskvoice-activity-detection — a per-frame speech/non-speech probability
Runtimeonnxruntime-web directly — transformers.js has no VAD pipeline, so this page drives ONNX Runtime Web itself (isolated to this model's worker)
ArchitectureSmall stateful RNN (LSTM). Inputs: input [1×512] frame · state [2×1×128] LSTM state · sr (16000). Outputs: speech probability + next state.
Frame512 samples @ 16 kHz = 32 ms per decision; the LSTM state carries context across frames
Quantizationfp32 ONNX (the whole model is ~2 MB — no quantisation needed)
BackendWebAssembly via onnxruntime-web (single-threaded; runs anywhere)
Download~2 MB, cached in Cache Storage after first load (auto-inits next visit)
LicenseMIT
Web APIsWebAssembly, Web Workers, Web Audio, MediaRecorder, Cache Storage, Canvas — all Baseline widely available

Run it

Pick a sample, drop an audio file, or record from your mic, then hit Analyse. Silero walks the clip frame-by-frame and marks where speech is.

Drop an audio file here, or click to choose

The bundled Speech clip alone (not the other samples) is a JFK inaugural excerpt (20 January 1961), spoken by John F. Kennedy — public domain, U.S. government work · JFK audio source and attribution record.

0.50
100

See inside

The VAD doesn't return "speech: yes/no" — it returns a probability for every 32 ms frame. Everything else (segments, trimming, live meters) is what you do with that curve. Below is the real per-frame probability the model produced, with your decision threshold (dashed) and the frames above it shaded. Under it, the speech segments after a threshold + hysteresis pass (a frame must stay quiet for the "min silence" you set before a segment ends).

Detected speech segments (green = speech):

Analyse a clip to see its probability curve and speech segments.

Why it matters

VAD is infrastructure. It's cheap, it's fast, and it makes everything downstream better — which is exactly why it belongs on-device: no audio needs to leave the browser just to decide whether it's worth processing.

Use cases

Basics

Is there speech in this clip? Get a yes/no plus the probability curve behind it.

Practical

Trim the silence and split a clip into speech segments you can export.

Wild

A live mic talk-meter — watch the speech probability move in real time as you talk.

Multi-model

VAD → Whisper: transcribe only the speech segments, skipping the silence.

How the API works

There's no pipeline() for VAD — you drive onnxruntime-web directly. It's a stateful loop: feed 512-sample frames, carry the LSTM state forward, read the probability out.

import * as ort from "onnxruntime-web";

const session = await ort.InferenceSession.create(
  "https://huggingface.co/onnx-community/silero-vad/resolve/main/onnx/model.onnx",
  { executionProviders: ["wasm"] },
);

let state = new ort.Tensor("float32", new Float32Array(2 * 1 * 128), [2, 1, 128]);
const sr = new ort.Tensor("int64", [16000n], []);          // scalar

for (const frame of frames512) {                            // Float32Array, 512 samples @ 16 kHz
  const input = new ort.Tensor("float32", frame, [1, 512]);
  const { output, stateN } = await session.run({ input, state, sr });
  const speechProb = output.data[0];                        // 0 … 1 for this 32 ms frame
  state = stateN;                                           // carry the recurrent state forward
}

This page decodes files / mic recordings to 16 kHz mono with an OfflineAudioContext, runs the frame loop inside a Web Worker so the UI never janks, caches the ~2 MB model in Cache Storage (so a return visit auto-initialises with no download), and turns the raw curve into segments with a threshold + hysteresis pass.

References