VADConfig

Defined in: src/core/types/config.ts:380

Configuration for local voice activity detection.

Remarks

When enabled, a local VAD model (Silero by default, running on ONNX Runtime) scores the microphone audio directly, independent of the STT provider. This gives the pipeline:

• Provider-independent barge-in — playback is interrupted as soon as sustained local speech is detected, without waiting for the STT provider to return text.
• False-barge-in resistance — a higher bargeInThreshold applies while the agent is speaking, so TTS echo leaking into the microphone has to clear a higher bar; VADConfig.minSpeechDurationMs debounces transient noise.
• Configurable end-of-turn sensitivity — VADConfig.silenceDurationMs controls how much silence ends a speech segment (vad.speechEnd).

Requires the onnxruntime-web (browser) or onnxruntime-node (server) optional peer dependency unless a custom VADConfig.engine is supplied. Only available when the input provider is separate from the STT provider (multi-role inputs like NativeSTT manage their own microphone, so the SDK never sees the raw audio).

Example

const agent = new CompositeVoice({
  providers: [...],
  vad: {
    enabled: true,
    modelUrl: '/models/silero_vad_v5.onnx',
    bargeIn: true,
    silenceDurationMs: 600,
  },
});
agent.on('vad.speechStart', ({ probability }) => showSpeakingIndicator());
agent.on('vad.speechEnd', ({ durationMs }) => hideSpeakingIndicator());

See

• CompositeVoiceConfig.vad for where this is used
• TurnTakingConfig for the complementary echo-avoidance system

Properties