What is endpointing?
Endpointing is the decision that a speaker has finished and the system may respond. It is taken from silence duration alongside acoustic and linguistic cues, usually within a few hundred milliseconds. Because every other component waits on that decision, endpointing determines how responsive a call feels more than raw model speed does.
Also called: End of speech detection, End-of-utterance detection
Why responsiveness is settled before any model is invoked
Teams chasing a faster call almost always examine the model first, and the delay is usually elsewhere. A system that waits most of a second of silence before deciding the caller has stopped has already spent the pause a caller will tolerate, however quickly the response is then generated. Endpointing too eagerly is worse, since the agent then interrupts a caller who paused mid-sentence to recall an account number. The setting is a judgement about the specific task, and a threshold that suits a yes or no answer will fail on a long address.
- Endpointing delay is added to every single turn of the conversation.
- Callers experience the pause before a response as the system's speed.
- An early endpoint truncates the caller and loses the end of the utterance.
- The right threshold differs per field, so one global setting fails somewhere.
How endpointing works
- Voice activity detection separates speech frames from silence on the caller channel.
- A silence threshold is applied, measured from the last detected speech frame.
- Acoustic cues such as a falling pitch contour inform the decision.
- The partial transcript is examined for whether the utterance is syntactically complete.
- Thresholds are configured per field, with longer waits where digits are expected.
- The decision is logged with its timing, so latency can be attributed afterwards.
How endpointing is evaluated
Endpointing is evaluated on two errors that pull against each other, so both are measured together. Examine the following on recorded turns:
- median delay from end of speech to the endpoint decision
- rate of premature endpoints that truncate the caller
- rate of late endpoints where the caller waits in silence
- accuracy on utterances containing a mid-sentence pause
- endpointing behaviour per field type, including digit strings
Limitations, risks and common failure modes
- A threshold tuned on short answers truncates callers reciting an address.
- Thinking pauses before a number are treated as the end of the turn.
- Background speech keeps voice detection active, so the endpoint never arrives.
- Tightening the threshold for speed increases truncation that nobody measures.