@guidekit/vad

Voice Activity Detection package for the GuideKit SDK. Wraps the Silero VAD ONNX model to detect when a user is speaking, enabling half-duplex voice interactions with barge-in detection.

Installation

npm install @guidekit/vad

This package is an optional peer dependency of @guidekit/core. It is automatically used when voice mode is enabled.

How It Works

@guidekit/vad loads the Silero VAD model via ONNX Runtime Web and runs inference on audio frames from the microphone. It emits speech_start and speech_end events that the core engine uses to trigger STT and manage barge-in.

Usage with GuideKit

When using @guidekit/react or @guidekit/vanilla, simply install this package and enable voice mode:

<GuideKitProvider
  tokenEndpoint="/api/guidekit/token"
  agent={{ name: 'Guide' }}
  options={{ mode: 'voice' }}
>
  {children}
</GuideKitProvider>

The SDK detects the package automatically. If it is not installed and voice mode is requested, the SDK emits a VAD_PACKAGE_MISSING error and falls back to text-only mode.

Standalone Usage

import { createVAD } from '@guidekit/vad';

const vad = await createVAD({
  onSpeechStart: () => console.log('Speaking...'),
  onSpeechEnd: (audio) => console.log('Done, audio length:', audio.length),
});

// Start processing audio from a MediaStream
await vad.start(mediaStream);

// Stop
vad.destroy();

Documentation

• Full documentation: guidekit-docs.vercel.app/docs/voice
• Main README

License

MIT