Endpointing
The decision process that determines when enough end-of-speech evidence exists to finalize an utterance and send it to downstream recognition.
Different from VAD
VAD estimates whether speech is present in a frame. Endpointing decides whether an utterance has finished.
A single silent frame is not necessarily an endpoint because normal speech contains pauses inside words, phrases and sentences.
Latency vs Truncation
Aggressive endpointing reduces user-visible delay but can cut an utterance too early. Conservative endpointing preserves context but adds latency.
System Perspective
The decision also changes queue and worker behavior. More small segments can create additional scheduling and inference overhead.
I discuss that relationship in ASR Capacity Engineering.