The decision
Our models do one job: at every pause, they decide whether the speaker has finished or is just pausing.
Get this wrong one way and the agent cuts people off. Get it wrong the other way and it waits too long to reply. Both feel broken to the person on the call, so a detector has to be judged on both at once.
One call, step by step
- Caller
I'd like to move my appointment to
Pause. The detector keeps listening. - Caller
Thursday afternoon, if you have anything.
Finished. End of turn, the agent replies. - Agent
Thursday at two is open. Shall I
The caller starts talking over the agent. - Caller
Two is perfect.
Interruption. The agent stops.
An example, not model output. The first pause is mid-sentence, so the detector holds. The second is the end of the turn, so the agent can reply. When the caller cuts in, the detector marks an interruption and the agent stops speaking.
Where it sits
A voice agent has four jobs: hear the words, know when the speaker is done, decide what to say, and speak. Our models do the second one. They work with any speech recognizer, any LLM and any voice.
How we report it
Three numbers matter together: recall, the share of true events the detector catches; false-positive rate, how often it fires when it should not; and latency, how long after the event it decides. We quote latency as the median and show the p10 and p90 where we have them.
Our current numbers are in the TurnBench post and on the turn-1 page.