Silero VAD: browser setup for a voice assistant

This tutorial, uploaded April 7, 2025, demonstrates Silero VAD in a Next.js browser interface using JavaScript and the @ricky0123/vad-web package. The source gives no recording date or numbered package version. Its assistant example uses Llama Scout, so the setup reflects that period rather than a current compatibility guide.

The first demo changes text and animates circles around a microphone icon when speech starts. The presenter explains two callbacks: onSpeechStart triggers application logic, while onSpeechEnd supplies an audio clip for further processing. Speech detection runs in the browser; transcription is a separate step. The presenter describes the model as roughly 2 MB and MIT licensed, and cites its README's claim that a single CPU processes a 30-plus-millisecond audio chunk in under one millisecond.

The fuller assistant sends captured speech to Whisper through Groq, passes the text to Llama Scout through the same provider, and uses Cartesia to generate spoken replies. This combines on-device detection with hosted services rather than a fully offline assistant. The presenter reports that the package helped prevent speaker audio from feeding back into transcription, but this is an observation from his setup, not a universal echo-cancellation guarantee. He also briefly discusses a Qwen multimodal alternative without identifying a specific version.