endpointing is how a voice agent decides that the caller has finished speaking. once the endpoint closes, the system treats the transcript as a complete turn and prepares a reply.
the difficult part is telling a finished turn from a pause. callers stop briefly to think, breathe, find a number, or change what they were about to say.
how does endpointing work?
in silence-based endpointing, the system detects a change from speech to silence. it waits for a set period, then closes the turn if the caller does not continue.
a fixed delay treats every pause the same way. other systems can also look at the partial transcript or use a turn-detection model. "the account number is" sounds unfinished so a longer wait makes sense even though the microphone has gone quiet.
in a silence-based system, a shorter delay lets replies begin sooner but increases the risk of cutting off a hesitation. a longer delay gives the caller more time to continue, but adds silence after a complete answer.
livekit's endpointing controls place minimum and maximum limits around this wait. the delay can remain fixed or change with the pause patterns heard during the call.
endpointing example
a caller tells a billing voice agent, "the invoice number is... hang on... four eight two zero nine three."
the caller may go quiet twice before reading the number. voice activity detection can mark both gaps as silence but closing the turn after either one would be a mistake. the sentence is unfinished and the caller has signalled that more information is coming.
after "three," the caller stays quiet. now the endpoint can close with the full invoice number in the transcript.
a short reply behaves differently. if the caller answers "yes," holding the turn open for the same length of time adds delay without recovering any missing information. endpointing has to handle both kinds of reply in the same call.
endpointing vs vad
voice activity detection identifies speech and silence. endpointing uses those boundaries to decide whether the caller's turn is complete.
difference | vad | endpointing |
|---|---|---|
question | does this audio contain speech? | has the caller finished their turn? |
input | short audio chunks | speech boundaries, pause length, and sometimes transcript context |
output | speech or non-speech | turn open or turn complete |
common error | misses quiet speech or reacts to noise | closes during a hesitation or waits too long |
vad may detect the pause after "hang on" exactly as it happened then an endpointing delay closes the turn too soon. the opposite can happen when vad misses a quiet word and gives endpointing the wrong boundary.
why does endpointing matter for voice agents?
callers do not always finish a request in one breath. someone might say: "send it to my office... actually, use my home address."
if the voice agent starts answering during that pause, it misses the correction. if it waits too long after the caller finishes, they may repeat themselves because they think nobody heard them.
in rumik's livekit voice-agent setup, livekit handles turn-taking and rumik.TTS() speaks the reply. endpointing determines which words belong to the caller's request. the rumik voice can start quickly, but the reply still needs to account for what the caller finished saying.
a caller looking up a number needs more time than someone answering "yes." good endpointing leaves room for that hesitation without making every short answer wait through the same long silence.