End-of-Turn Detection

Step 1It is a prediction about completion, not a measurement of silence.

Three things in a voice stack are routinely conflated, and keeping them apart is most of the value of the term.

• Voice activity detection (VAD) answers “is there speech in this frame?” It is cheap, acoustic, and runs continuously. It knows nothing about whether a sentence is finished.
• Endpointing is the policy layer that turns VAD output into a decision: typically, “speech stopped and has stayed stopped for N milliseconds, so the turn is over.”
• End-of-turn detection is the prediction that the utterance is complete — that the speaker has said what they meant to say and is yielding the floor. Completion is a property of the content and the prosody, not of the energy level.

The distinction has a canonical example. A caller says “my account number is four seven” and pauses. VAD correctly reports silence. An endpointer with a 500 ms timer correctly reports that silence has lasted 500 ms. Both are right, and the agent that interrupts here is wrong, because the utterance is obviously unfinished to any listener who heard the words.

Step 2No silence threshold separates the two distributions.

The reason a timer cannot be tuned into correctness is statistical rather than engineering. Pauses within a turn and gaps at the end of a turn are drawn from two distributions that overlap heavily. Any threshold you pick cuts through both of them, so every setting buys a reduction in one error by paying for an increase in the other.

# The timer's two errors, and why they trade
threshold too low    # fires inside a mid-turn pause  -> false cutoff
threshold too high   # waits out a finished utterance -> dead air, every turn

# And the second cost is unconditional:
added_latency = threshold          # paid on EVERY turn, not just the ambiguous ones
cutoff_rate   = P(pause > threshold | mid-turn)

The overlap also moves with the speaker, which is what defeats a per-deployment tuning exercise. Second-language speakers, older callers, people reading a number off a screen and anybody thinking while they talk all sit further right on the pause distribution than a fluent native speaker finishing a sentence. One vendor write-up puts mean pauses for Mandarin–English bilinguals around 686 ms against roughly 546 ms for native English speakers — a figure we have not been able to trace to a primary dataset, but directionally the familiar one. A threshold tuned on your own team is a threshold tuned on the fastest, most fluent tenth of your callers.

The consequence for multilingual agents is sharper than it looks: one global timeout encodes a different interruption rate for each language group you serve, and nothing in your dashboard will break it out that way.

Step 3The 2026 shape: the biggest line is now a model.

The market answer is to predict completion directly from the signal rather than infer it from duration. A model-based end-of-turn detector consumes the audio, the partial transcript, or both, and emits a probability that the speaker has yielded — which can fire before the silence timer would have, on an utterance that sounds finished, and hold the floor open through a long pause on one that does not.

The architectural move of 2026 was to collapse this into the recognition model itself. Deepgram's Flux, launched that year, performs transcription and turn detection in one model and so needs no separate VAD-plus-endpointing stage; Deepgram documents end-of-turn decisions at roughly 260 ms, and its multilingual variant at under 400 ms across ten languages. Those are vendor figures measured on the vendor's own path, which is the usual caveat, but the structural claim underneath them is checkable and matters more: fusing the two removes a serial hop, and a serial hop is the thing the latency budget is made of.

Two practical notes survive the vendor churn. Model-based detection is normally run alongside acoustic VAD rather than instead of it — VAD is near-free and still the fastest way to know that the caller has started talking, which is what barge-in needs. And detectors expose a decision threshold, so adopting one does not remove the operating-point choice below; it gives you a better curve to pick a point on.

Step 4Pick an operating point, and never report one side of it.

There is no standard for an acceptable false-interruption rate, so the number is yours to set and yours to defend. Set it per use case: an agent taking a payment card or a booking reference should hold the floor through long pauses and accept the added latency, because an interruption there costs a restart of the whole value; an agent handling short confirmations can fire early because the cost of a wrong cut is one repeated word.

The reporting discipline is the part teams skip. A latency improvement that was bought by lowering the threshold is not an improvement, it is a transfer, and it will show up weeks later as abandoned calls that nobody connects to the change.

Finally, know what happens when the event disappears. A full-duplex model that listens while it speaks never emits an end-of-turn event at all, and every downstream consumer that was silently subscribed to it — tool dispatch, trace spans, guardrail checkpoints, handoff logic — loses its trigger. That is a change of architecture rather than of latency, and it is the subject of full-duplex speech.