Endpointing is the moment the agent decides 'the caller is done talking, my turn now.' Set the threshold too short and the agent interrupts every pause. Set it too long and the call drags. The right threshold is conversational, context-dependent, and per-vertical.
Naive endpointing uses a fixed silence timeout, usually 600-1200 ms, and lives with both failure modes. Better endpointing uses VAD output combined with prosody, syntactic completeness, and intent signals. Best is a learned model that predicts end-of-turn from the partial transcript and the audio jointly.
Endpointing failures are the second-biggest complaint customers have about voice agents (after latency). A vendor that has not invested in this layer ships an agent that feels rude, even if every other layer is excellent.
Vorel ships per-vertical endpointing overrides, different verticals get different silence thresholds and prosody weights by default.