N.O.V.A orbit and star emblemN.O.V.A v2
N.O.V.A v2 / documentation

Voice Lane

The architecture behind every voice command, end to end.

Pipeline

Mic stream→Wake word→STT (VAD + Whisper)→Rule match→LLM router→Tool + reply

Stages

  • Mic stream — always running, low cost.
  • Wake word — openWakeWord decides when to wake the heavier models.
  • Speech-to-text — Silero VAD finds where the command ends, faster-whisper transcribes it.
  • Rule match — known phrases resolve instantly without touching the LLM.
  • LLM router — anything unmatched goes to the local model, which returns a tool call as JSON.
  • Tool execution + reply — Python runs the tool; Piper speaks the result. A cloned voice using XTTS-v2 is being trialled as an alternative.

See Voice Commands for the tools themselves, and Recorder Lane for the other half of the system.

PREVTroubleshootingNEXTRecorder Lane →