+ Document & Voice AI · Flagship
Voice AI Receptionist
Low-latency voice agent that books dental appointments: browser VAD, Groq Whisper, a tool-calling LLM on SQLite and per-sentence TTS with barge-in.
~1 sp50 from end of speech to first reply
Overview
A full-duplex voice receptionist for a fictional dental clinic. It listens through the browser, transcribes with Groq Whisper and reasons with a streaming tool-calling LLM that checks availability and books, reschedules or cancels appointments in SQLite. Replies are spoken sentence by sentence as they stream, and callers can interrupt at any time. Every turn records per-stage latency, shown as badges in the UI and summarised by a benchmark script.
Highlights
- Measured p50 of 980 ms from end of speech to first reply sentence (STT p50 326 ms, LLM first token p50 690 ms)
- Full pipeline: VAD, WebSocket PCM streaming, STT, streaming LLM with 6 SQLite tools, per-sentence TTS and barge-in
- Confirmation gate that blocks bookings unless the exact details were read back first
- Benchmark reporting p50/p95 per stage; graceful fallback from server TTS to browser speech
- 103 offline tests with injected fakes for STT, LLM and TTS