What is Endpointing (Turn-Taking)?
Endpointing is how a voice agent decides that the caller has finished speaking and it is now its turn to respond. Getting it wrong means either cutting people off or leaving awkward silences.
Also known as: end-of-turn detection, turn-taking, voice activity detection
How it works
The simplest method waits for a fixed stretch of silence, say a few hundred milliseconds, then answers. Better systems combine voice activity detection with the content of the sentence: 'my number is nine eight' is clearly unfinished even if the caller pauses, while 'yes please' is complete almost instantly. Some also learn per-caller pacing.
The trade-off
Endpointing sits directly on top of latency. Waiting longer is safer against interruptions but adds dead air to every turn; answering sooner feels quick but talks over people who pause to think. Dictating a phone number, a PIN or an address is where fixed silence thresholds break down hardest.
Example
A caller says 'I want to book for... um... Saturday morning'. A pure silence timer replies after the first pause and asks 'Which day?'. A content-aware endpointer waits, because 'book for' ended mid-thought, and answers only after 'morning'.
Where this shows up at Vaami
Vaami targets sub-300ms end-to-end response latency in production, which includes the end-of-turn decision.
See it on a real call.
Hear how a Vaami voice agent handles this on a live conversation in your language and your use case.