Architecture
The full path of a single turn.
Six stages, wired by the router. Barge-in loops the last stage back to the first.
Wake loop
openWakeWord · barge-in
Router
voice-agent-proxy
STT
faster-whisper
Agent chain
A2A + LLM · MCP · memory
TTS chain
minimax · 11labs · openai
Audio
cached or streamed
↻ barge-in loops back to the wake loop · every stage reports its provider in the response so you know exactly what served the turn.
microphone/wake loop (openWakeWord, barge-in)
-> voice-agent-proxy router
-> faster-whisper /v1/audio/transcriptions
-> chat backend chain: A2A peer and/or OpenAI-compatible LLM
(+ MCP tools, per-agent conversation memory)
-> TTS chain: MiniMax | ElevenLabs | OpenAI-compatible
-> audio response (cached file or streamed as it synthesizes)