Silero VAD

The Silero VAD plugin provides AI-powered voice activity detection for barge-in (interruption) detection. It uses the Silero VAD machine learning model running via ONNX Runtime WebAssembly to distinguish real speech from background noise, coughs, and other non-speech sounds.

Installation

pnpm add @ariontalk/plugin-silero-vad

The package exports SileroVadDetector and the SileroVadOptions type.

Usage with the widget

Register the Silero VAD plugin alongside other barge-in strategies by setting the bargeInPlugins property on the widget element:

<ariontalk-widget id="widget" settings></ariontalk-widget>

<script type="module">
  import '@ariontalk/widget';
  import { EnergyBargeInDetector } from '@ariontalk/core';
  import { SileroVadDetector } from '@ariontalk/plugin-silero-vad';

  document.querySelector('#widget').bargeInPlugins = [
    {
      id: 'energy',
      label: 'Energy',
      tooltip: 'Interrupt by speaking — uses mic energy detection',
      create: () => new EnergyBargeInDetector(),
    },
    {
      id: 'silero-vad',
      label: 'Smart VAD',
      tooltip: 'AI-powered speech detection — more accurate, fewer false triggers',
      create: () => new SileroVadDetector({ onnxWASMBasePath: '/' }),
    },
  ];
</script>

Configuration options

Pass a SileroVadOptions object to the constructor to customize detection behavior.

Option	Type	Default	Description
baseAssetPath	string	'./'	Base path where vad.worklet.bundle.min.js and silero_vad.onnx are served.
onnxWASMBasePath	string	'./'	Base path where ONNX Runtime WebAssembly files are served.
positiveSpeechThreshold	number	0.7	Probability threshold for speech detection (0-1). Higher values reduce false positives but are slower to trigger.
negativeSpeechThreshold	number	0.55	Probability below which speech is considered absent (0-1). Should be lower than positiveSpeechThreshold (Silero recommends ~0.15 less).
minSpeechMs	number	500	Minimum sustained speech duration in milliseconds before triggering barge-in. Shorter sounds (coughs, thumps) are discarded.
redemptionMs	number	1400	Grace period in milliseconds after speech drops below negativeSpeechThreshold before considering speech ended. Bridges brief pauses mid-sentence.

Example with custom options

create: () => new SileroVadDetector({
  onnxWASMBasePath: '/assets/wasm/',
  baseAssetPath: '/assets/vad/',
  positiveSpeechThreshold: 0.8,
  negativeSpeechThreshold: 0.6,
  minSpeechMs: 400,
  redemptionMs: 1000,
})

Asset serving

Silero VAD requires several files to be served from your web server:

• ONNX Runtime WASM files — The WebAssembly runtime for running the ML model. Set their location with onnxWASMBasePath.
• VAD worklet and model — The vad.worklet.bundle.min.js audio worklet and silero_vad.onnx model file. Set their location with baseAssetPath.

These files come from the @ricky0123/vad-web package (a dependency of the plugin). Copy them to a publicly served directory or configure your bundler to serve them from node_modules.

The model and WASM files are loaded lazily — they are only downloaded when the user starts a session with Smart VAD selected, not at page load.

Energy vs. Silero VAD comparison

Aspect	Energy (built-in)	Silero VAD
Detection method	RMS volume threshold	ML-based speech probability
Accuracy	Good — may trigger on loud non-speech sounds	High — distinguishes speech from noise
False positives	More likely (keyboard typing, music, etc.)	Rare — trained to detect human speech
CPU usage	Minimal (simple math on audio samples)	Moderate (ONNX model inference via WASM)
Bundle size	Zero (included in @ariontalk/core)	~2 MB (ONNX Runtime WASM + Silero model)
Setup	None — works out of the box	Requires serving WASM and ONNX assets
Dependencies	None	@ricky0123/vad-web, onnxruntime-web
Latency to trigger	~250ms of sustained volume	~500ms of sustained speech (configurable)