Turn detection
In one sentence
Turn detection is working out whose turn it is to speak, and semantic turn detection does it by judging whether the sentence sounds finished rather than by timing the silence after it.
Not to be confused with Endpointing, Voice activity detection, or Barge-in.
Definition
Turn detection is a voice agent deciding whether you have handed the conversation back to it or are still mid-thought.
The semantic version reads what you have said so far and asks whether it sounds like a finished sentence. The older approach only counts how long you have been quiet.
A silence timer is asking something acoustic: has nobody spoken for N milliseconds. The semantic version asks something linguistic: is what I have heard so far a complete thought, and is it inviting an answer. The second question is the better one, because the length of a pause is not what governs human turn-taking. People read syntax, prosody and content to predict where a sentence is heading, and that prediction is how conversational gaps come in around 200 milliseconds without everybody colliding.
How it is built
- A small classifier reads the partial transcript as it arrives and returns the odds that the turn has ended.
- That probability moves the silence threshold up or down as the sentence unfolds. High confidence the turn is done permits a short wait. Low confidence extends it.
- Some implementations also read acoustic prosody, since falling pitch signals completion while sustained or rising pitch signals that more is coming.
What it fixes that a timer cannot
- "My number is oh seven nine..." scores as plainly unfinished, so the system waits through the pause rather than answering into it.
- "That's all, thanks" scores as plainly finished, so the system can reply almost at once instead of sitting out the full threshold.
- The result is faster in one case and more patient in the other, from a single mechanism. This is the rare change that improves both sides of a tradeoff instead of moving a slider along it.
What it does not fix
- It puts a model inference inside the turn-taking path. The models are small and the cost is modest, but it is not free.
- It is tied to one language. Train the classifier on English completion patterns and it will not carry across to another language, and code-switching defeats it.
- Genuinely ambiguous utterances stay ambiguous. Trailing off is trailing off, and no classifier resolves what the speaker has not decided.
Two ideas that sit next to it
- Full-duplex conversation, where both parties can talk at once and the system handles the overlap instead of enforcing strict turns.
- Backchanneling, the small "mm-hm" and "right" noises a listener makes. Some systems now produce them, and that changes the dynamics of a turn considerably.
Common misconception
That semantic turn detection is a marginal refinement. It is closer to a change of category in how a conversation feels, because it removes the forced choice between fast and patient. Systems using it are noticeably harder to interrupt at the wrong moment and noticeably quicker to answer once a question is plainly finished.
Why it matters commercially
When somebody says a voice agent felt human, or felt like a machine, turn-taking is what they are describing. They almost never have the vocabulary for it and they certainly never blame a threshold, but this is the mechanism producing the feeling. It also separates platforms in a real way right now, at a point where recognition and synthesis quality are converging.
In voice specifically
In text, a half-written message is invisible until you send it. In speech, everything you have said so far is already out, so the system has to decide in real time whether your pause means you are thinking or means you are done.
Visual
Silence timers against sentence understanding
Humans do not time silences. They predict endings.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
The average gap between turns in human conversation is around 200 milliseconds, and people achieve it by predicting where a sentence is going rather than by measuring silence. That is the behavior semantic turn detection approximates.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - A reference point rather than a target. It matters here because it is evidence that the human mechanism is prediction, not timing, which is the whole argument for the semantic approach.
The baseline it improves on is a silence-based endpointing threshold of 150 to 300 milliseconds applied uniformly to every turn.
150 to 300 msindustry rangeCommon production budgeting convention, 2026 - An engineering convention rather than a research finding. The tradeoff that comes with applying one number everywhere is set out on the endpointing page.
Barge-in is expected to cut synthesis within roughly 60 milliseconds of a detected interruption, and Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond. Together those figures indicate the acoustic machinery is already fast enough, and the remaining gains sit in the intelligence layer.
60 ms and under 1 msindustry rangeIndustry working threshold and Silero VAD published benchmark, 2026 - Two figures of different status quoted for one conclusion. The 60 millisecond budget is a working convention; the Silero measurement is published by the project itself. The conclusion drawn from the pair is directional.
No established independent benchmark for turn detection accuracy exists, and there is no common test set for comparing semantic turn detection implementations. Vendor claims in this area cannot currently be verified by a third party.
-no reliable figureSay this plainly. Few capabilities in the category are marketed as hard as this one and measured as little. Ask for a live demo rather than a figure, since pausing mid-sentence settles it in seconds.
Examples
In practice
A team swaps a fixed 300 millisecond endpointer for a semantic turn detector that moves the threshold between 100 and 900 milliseconds according to completion probability. Median time to first audio falls, because most turns are unambiguously finished and now resolve sooner. Complaints about interruption fall as well, because the ambiguous turns get more patience. Two metrics that normally trade against each other both improve from one change.
The everyday version
The difference between an agent that can tell you have finished your sentence and one that is only counting the silence. The first lets you pause while you hunt for the word. The second jumps in.
Usage
Who says it
- Voice AI engineers, and platform vendors who increasingly put it forward as a headline differentiator.
- Buyer-facing material more and more often, though usually described in outcome terms as natural conversation or interruption handling rather than by mechanism.
- Buyers themselves, rarely, and almost never in mechanism terms, which makes it a good thing to explain plainly if you can.
Where it turns up
- In platform feature comparisons and technical documentation.
- In demo scripts, because pausing mid-sentence shows the difference live. It is one of the few architectural choices with an effect a buyer can feel inside thirty seconds.
Common misuse
- Calling any tuned endpointing semantic turn detection. If no model is reading the transcript, it is a silence timer with adaptive settings.
- Claiming language coverage without evidence. Completion patterns differ by language and code-switching is still handled poorly.
- Implying it removes inappropriate interruption altogether. It reduces it. Ambiguity remains ambiguous.
Questions people ask
What is the difference between turn detection and endpointing?
Endpointing is acoustic. It counts milliseconds of quiet and closes the turn once it reaches N. Semantic turn detection is linguistic. It reads the words heard so far and judges whether they add up to a finished thought. The test to apply to any vendor claim is simple: is a model reading the running transcript. If nothing is, what you have is a silence timer with adaptive settings.
How does semantic turn detection work?
A small classifier reads the partial transcript as it arrives and outputs a probability that the turn is finished. That probability moves the silence threshold up or down in real time, so a plainly complete sentence gets answered quickly and an unfinished one gets more room. Some implementations also read prosody, because falling pitch signals completion and rising pitch signals that more is coming.
Does semantic turn detection make a voice agent faster or more patient?
Both, in the situations where each is right. Most turns are unambiguously finished, so they resolve sooner than a fixed threshold would allow, which pulls the median response time down. The ambiguous ones get more patience, which reduces interruption complaints. Improving both sides of a tradeoff from one change is rare enough to be worth checking for yourself.
What are the limits of semantic turn detection?
It adds a model inference into the turn-taking path, which is a modest but real cost. It is language-dependent, so an English classifier does not transfer and code-switching is handled poorly. Genuinely ambiguous speech stays ambiguous. And there is no independent benchmark or common test set, so no vendor claim about accuracy here can be checked by a third party today.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.