Audio and Turn Taking

The runtime reactor processes audio frames, model events, tool events, and timers in one event loop.

Core Pipeline

1. Ingest and normalize incoming audio
2. Run denoise/VAD stages
3. Stream STT updates
4. Apply turn-complete rules (VAD + smart-turn + transcript gate)
5. Trigger LLM and stream TTS response

Interruption Handling

When user speech arrives during assistant output:

• Runtime can interrupt immediately
• Or defer interruption until word-count / timeout gates are met
• Fallback timeout prevents stalled turns

Turn Safety Guards

• Smart-turn prediction does not commit by itself
• Runtime waits for non-empty transcript before advancing
• Late transcripts are dropped when timeout guard has already reset state

Idle Policy

Runtime tracks inactivity and can emit re-engagement prompts before hangup.