Jun 28, 2026 路 1 min read
Shipping a real-time turn detector
How the end of utterance of user gets detected via semantic and acoustic understanding.
The naive approach waits for silence. That is already too late.
Where the milliseconds go
Every hop between the microphone and the model adds delay. The trick is to predict the boundary from the shape of the audio, not to wait for a gap that has already cost time.
What it changes
- Turns end sooner, so the agent replies sooner.
- No larger model on the hot path.
- The same pipeline runs on device.
The fastest turn is the one you saw coming.
