Voice Agent Architecture — Hear, Think, Speak

How a real-time voice agent is wired — VAD and STT, orchestrator and LLM with tools, streaming TTS, barge-in, latency budgets, and phone vs web deployment patterns.

1 · Listen (Perception)

Transport: WebRTC, SIP, websocket audio

2 · Think (Cognition)

Dialog state + session memory

3 · Speak (Expression)

TTS streams audio — don’t wait for full text

Orchestrator

Owns turn state machine