Endpointing is the task of deciding when a speaker has finished an utterance. In streaming speech-to-text it detects the end of speech, the trigger that closes a segment and finalizes its transcript. In a voice agent it is the signal that the person has stopped talking, so the system can stop listening and respond.
How endpointing works
Endpointing builds on voice activity detection, which flags speech versus silence frame by frame. The endpointer watches that stream and waits for a configured gap of silence, often between 500 and 800 milliseconds, before declaring the utterance finished.
Simple endpointers use that silence timer alone. Stronger ones weigh acoustic cues like falling pitch, and sometimes linguistic ones, so a mid-sentence pause is not mistaken for the end of a thought.
Newer end-to-end systems fold the endpoint decision into the recognizer itself, predicting an end-of-speech token instead of waiting on a fixed timer. In streaming APIs that event usually flips interim results into a final, stable transcript.
The endpointing latency tradeoff
The silence threshold is endpointing's central tuning knob, and it forces a tradeoff you cannot escape. Set it short and the agent feels snappy but cuts people off when they pause to think or read a number aloud. Set it long and the agent is patient but feels sluggish, because every reply waits out that gap.
Two things sit in tension: responsiveness and the false-endpoint rate. This is why endpointing sits so close to turn detection, and why the two are usually tuned alongside barge-in, which lets a user interrupt a talking agent.
How endpointing relates to turn detection
Endpointing and turn detection overlap and get used loosely, but the distinction is useful. Endpointing is the acoustic decision that a stretch of speech has ended so the recognizer can close a segment. Turn detection is the higher, conversation-level call about when the floor should change hands from person to agent.
A silence-based endpointer is often the crude first version of turn detection. A good voice agent replaces or augments it with a model that ends on completion, not on silence.
Why endpointing matters
In a voice agent, endpointing delay is a large and often overlooked share of perceived response time. A user does not feel model latency and endpointing latency as separate things; they feel the total gap before the agent speaks.
Trimming the silence timeout, or switching to a smarter endpointer that ends on linguistic completion rather than raw silence, can make an agent feel far more responsive with no change to the underlying model.
Frequently asked questions
What is a good endpointing silence timeout? An endpointing silence timeout usually defaults to somewhere between 500 and 800 milliseconds. Shorter feels faster but cuts people off mid-pause; longer feels laggy. The right value depends on whether callers speak in bursts, like reading a card number, or in full sentences.
What is the difference between endpointing and turn detection? Endpointing detects the end of speech acoustically. Turn detection decides when the conversational turn should pass, using words and prosody, not just silence, so it can hold through a thinking pause.
Why does my voice agent feel slow to respond? A long endpointing timeout is a common cause. The agent waits out the full silence gap before it even starts generating a reply, and that delay stacks on top of model latency.