Endpointing

What it is

Streaming speech recognition emits interim results while the caller speaks and a final result when it decides an utterance has ended. Endpointing is that decision, and in most engines it is driven by a silence duration: after N milliseconds without detected speech, the current words are committed as final and the agent's pipeline can proceed. Typical configurable ranges run from a few hundred milliseconds to a couple of seconds.

Endpointing is a component-level mechanism. End-of-turn detection is the broader decision, which may combine endpointing with linguistic cues and a separate turn model. Many simpler agents use endpointing alone, which is why they behave like a silence timer.

Why it matters when buying

The endpointing threshold is a direct lever on voice-to-voice latency and on false cut-offs, and it is rarely exposed to buyers. Telephone audio complicates it: background noise or comfort noise from a carrier can keep the detector "in speech" and delay the final result, while an aggressive threshold on a quiet line cuts callers off between the digits of a number. Vendors that tune it once, globally, are tuning for the average caller and nobody in particular.

What to ask

Ask whether the threshold is fixed or adaptive, and whether it changes when the agent expects a number or a spelled name. Ask what happens on a noisy line when silence is never fully detected. Ask whether the agent can force a final result when it has heard a complete answer, rather than waiting the full threshold.