Turn detection by language

On every hosted phone call, the agent decides a caller has finished a turn using two signals.

Two signals

Voice activity is the first: it tells the agent when the caller is making sound versus staying silent. It runs on every call, in every language.

The turn model is the second, and it does not run everywhere. Where it is available, it listens to the caller's words and intonation and judges whether the sentence sounds complete, rather than just whether the caller has gone quiet. That is what tells "I want to... let me check" apart from an actual end of turn, when a pure silence measurement cannot.

Two language classes

The turn model currently covers Arabic, Chinese, Dutch, English, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Portuguese, Spanish and Turkish. An agent in one of these languages gets both signals: voice activity narrows the window, and the model decides whether the sentence is actually finished.

Every other language — including Sinhala, Tamil and Estonian — runs on voice activity alone. With no model judging intent, the length of silence itself is the whole decision, so the hold that follows it matters much more.

How long the agent waits

After the caller stops speaking, the agent holds for a moment before treating the turn as finished. The interruption_sensitivity presets on the agent's voice stack move that hold, and by a different amount depending on the language class. In a turn-model language there are two waits: the shorter one when the model judges the caller finished, and the longer one when it judges them probably not finished, after which the turn ends anyway:

These are the phone-call values, for inbound and outbound calls alike. A phone line is narrowband, and on it the model too often judges an ordinary pause in the middle of a sentence (after the caller's name, before a number, while they leave a message) as the end of the turn, so a phone call holds longer than a browser test does. A language without the turn model needs the full hold to stand in for that judgment instead.

The interruption threshold — how long the caller must speak before they can cut the agent off — does not follow this split: it is one value for every language, 500 ms by default, 750 ms on low and 150 ms on high. See Interruptions and barge-in for that table and every other turn_taking setting.

The hold is a minimum

The hold is the shortest wait before a turn can end. It is not the time the caller waits for a reply. Two other things have to happen first, and the turn ends at whichever comes last:

• Voice activity reports the end of speech. That takes about 250 ms of silence, on every call. A hold shorter than that changes nothing.
• The final transcript arrives. The agent does not answer words it has not received as text. How long that takes depends on the speech-to-text provider, and on many stacks it is longer than the hold, so changing the preset there changes nothing a caller can hear.

The model and the voice also need their own time before the caller hears anything.

The agent editor shows this for each agent. Under Interruption sensitivity it lists the seconds behind all three choices for the agent's language (the shortest pause and, in a turn-model language, the longest wait when the caller may not be finished), and a timeline of the selected one: the caller stops talking, the shortest pause ends, and where that agent's recorded replies actually started.

Overrides

An explicit silence_hold_ms replaces the shorter hold above, in any language, including a turn-model one — set it directly instead of picking a preset. In a turn-model language the preset's longer wait still applies, and is raised to your value when your value is longer. An explicit min_speech_ms replaces the interruption threshold the same way, and is honoured on phone calls down to 150 ms.

Choosing a preset

In a language without the turn model, the hold is the entire end-of-turn decision, so it is worth tuning directly. In any language: if callers are being cut off mid-sentence, for example an agent that takes messages, move to low or raise silence_hold_ms; if the agent feels slow to respond, move to high or lower silence_hold_ms. That only helps down to the point where the final transcript arrives: if the editor's timeline shows the agent's replies starting well after the shortest pause, the wait is somewhere else, and a faster speech-to-text provider or model will do more than a shorter hold. Check the result on a phone call: browser test calls and the Simulations Lab do not apply these settings.

Whichever preset an agent runs, interruption itself is unaffected: barge-in is governed by the threshold above, not by the hold, so a caller who wants to cut in is never held up by how long the agent waits after they go quiet.