End-of-turn detection

What it is

A voice agent cannot see the caller's face, so it must infer from audio and words when a turn is over. The simplest method is a silence timer: after some hundreds of milliseconds without speech, assume the caller is done. Better systems combine acoustic cues (pitch falling, pace slowing) with what has been said (a complete sentence versus "my account number is..."), and some use a dedicated model trained to predict turn ends.

Every method has a floor. A fixed silence timer set short enough to feel responsive will cut off callers who pause to think, read a number or take a breath. Set long enough to avoid that, it adds its full length to every turn's latency.

Why it matters when buying

End-of-turn detection is where two of the most common caller complaints originate: "it kept interrupting me" and "it takes forever to answer". It also interacts with call type. Callers reading out a policy number, a date of birth or an address pause naturally between groups; an agent tuned for chatty callers will jump in.

What to ask

Ask how the agent decides a turn is over and whether the behaviour adapts to context (number entry, spelling, a hesitant caller). Ask what the wait is after the last word, at the median and the tail. In the demo, read out a phone number slowly with pauses between groups, and pause mid-sentence to think. The agent should wait in both cases and still answer promptly after a clear question.