What is turn detection?
It is why "uh, hold on, let me check" doesn't get treated as a caller hanging up their turn.
Turn detection
turn-taking
Turn detection is how a voice agent decides whether a caller has finished their turn or is still speaking, even through a pause. It is broader than watching for silence, because it can weigh what was said, not only how long the gap lasted.
A caller who says "my account number is... uh, hold on, let me check" has not finished. They have stopped talking. Those are different events, and a system that only measures silence cannot tell them apart.
Endpointing answers a narrower question: has speech stopped for long enough to count as a gap? Turn detection answers a wider one: given what was just said, is this actually the caller's turn to hand back? A trailing "um" or an unfinished clause is a strong signal the turn continues, even during a long pause.
Two layers: acoustic and semantic
Acoustic turn detection works from sound alone. A voice activity model, commonly Silero VAD, flags when speech starts and stops. A pause past a set length counts as the end of a turn. It runs fast. But it cannot separate a caller who is done from one who is thinking out loud.
Semantic turn detection adds a model that scores the content of the speech, and the gap. LiveKit ships this as its turn detector model. It predicts end of turn from the meaning of the speech, plus its acoustic properties, layered on top of Silero VAD rather than replacing it. OpenAI's Realtime API offers a comparable mode called Semantic VAD. It is tuned by a single setting named eagerness, not a raw millisecond threshold. Low eagerness gives a caller more room to keep talking. High eagerness cuts in sooner.
Neither approach replaces the other outright on current platforms. The VAD layer still triggers the fast reflex, such as stopping playback the instant a caller starts talking over the agent. Semantics sit above it, deciding whether a pause is a finished turn or a caller still forming the next sentence.
- Caller pausesSpeech stops for a moment.
- Acoustic checkVAD flags the gap length.
- Semantic checkWas that sentence finished?
- Turn resolvedContinue waiting, or hand back.
- Agent respondsOnly once the turn is actually over.
The semantic check is what current platforms add on top of a plain silence timer. It is why some agents let a caller trail off without losing the floor.
Common questions
01Is turn detection the same as endpointing?
No. Endpointing is the mechanism that measures a pause length and decides it counts as silence. Turn detection is the broader decision of whether the caller's turn has actually ended. It can weigh the content of the speech, not only the pause.
02What is semantic VAD?
Semantic VAD is voice activity detection that scores how likely a speaker has finished, based on what they said, beyond how long they paused. OpenAI's Realtime API ships this as a mode. It uses a single eagerness setting, not a raw silence threshold.
03Do most voice AI platforms use semantic turn detection by default?
The major platforms now offer it, but not always as the default. LiveKit's turn detector model layers semantic scoring on top of its bundled Silero VAD and is the framework's recommended setting. Confirm the default in your platform's current docs before launch, since defaults change between releases.
