In natural conversation, the gap between one speaker finishing and the next beginning averages around 200 milliseconds. Push past 500 and the listener starts to fill the silence. Push past 800 and they assume the line dropped.
Where the milliseconds go
A voice turn is not one model call. It is endpoint detection, transcription, retrieval against the Brain, reasoning, tool calls into the scheduling system, speech synthesis, and network transit. Any one of those can eat your entire budget.
- Endpointing: 90ms with a streaming VAD tuned per-vertical
- Transcription: streamed, first partial in 60ms
- Retrieval and reasoning: 180ms median with a warm cache
- Synthesis: first audio frame at 70ms
The trick is not speed, it is overlap
We start synthesizing the opening of a response while the reasoning for the rest of it is still running. Callers hear a natural start, and the remainder arrives before they notice a seam.