What is Endpointing?
Also called end of turn detection, end of utterance detection, EOU detection.
Endpointing is the decision a voice system makes about when a speaker has finished their turn and a response should begin. It combines silence timing with acoustic and linguistic cues to separate a natural mid-sentence pause from a genuine handoff. Getting it wrong produces either interruptions or noticeable dead air.
The simplest endpointer waits for a fixed period of silence, often a few hundred milliseconds, after speech stops. This is easy to implement and fails predictably: a speaker pausing to recall an address gets cut off, while a speaker who trails off leaves the system waiting. Fixed thresholds also interact badly with slow speech, disfluency, and languages with different rhythmic patterns.
Modern systems add semantic endpointing, where a model reads the partial transcript and estimates whether the utterance is syntactically and pragmatically complete. A trailing conjunction or a rising list intonation suggests more is coming, while a finished question suggests the floor is being handed over. Prosodic features such as pitch fall and final syllable lengthening carry similar information and can be scored directly from audio.
Endpointing sits on the critical path for perceived responsiveness. Any delay added here is added to every reply, so systems often start generating a response speculatively during the silence window and cancel if speech resumes. Some designs emit a short acknowledgment token to cover the gap, though overuse of filler makes the interaction feel scripted.
Evaluation typically tracks two error rates: premature cutoffs, which frustrate users mid-sentence, and late endpoints, which add latency. The right balance is task-dependent. Dictation tolerates longer waits, while a fast back-and-forth conversation needs an aggressive endpointer paired with reliable barge-in so mistakes are recoverable.
Key points
- Decides when the user's turn is over and a reply may start
- Fixed silence timers cut off thinking pauses
- Semantic and prosodic cues improve completion judgments
- Delay here is added to every single response
- Aggressive endpointing needs reliable barge-in as a safety net
In practice
A caller says their order number, pauses for nearly a second while reading the rest off a packing slip, then continues. A silence-only endpointer would have answered into that pause. The system instead scores the partial transcript, sees an incomplete number sequence, and extends its wait. When the caller finishes and the transcript parses as a complete identifier, the agent responds without any perceptible gap.