Endpointing

The naive implementation is a silence timer: once the VAD reports non-speech continuously for some threshold, declare the utterance over. This works acceptably for short factual replies and badly for everything else, because natural speech is full of pauses that are not endings — hesitation before a number, a breath mid-sentence, the gap while someone reads their card details off a screen.

Better endpointers combine several signals. Acoustic cues include falling pitch and lengthening on the final syllable, both reliable markers of a completed phrase in English. Lexical cues come from the streaming transcript: "my number is four one five" is clearly unfinished, while "that works for me" is clearly complete. Some systems predict endpoint probability directly from audio and partial text together.

Endpointing thresholds usually need to be context-dependent rather than global. When the agent has just asked for a long alphanumeric string, a longer threshold prevents truncation; when it has asked a yes-or-no question, a short one keeps the conversation snappy. Fixing a single value for the whole call means accepting the wrong trade-off in one direction or the other.