People rarely wait for silence before they answer. They hear the sentence closing: the pitch falls, the last word lands, and they start speaking within a fraction of a second.
We first wrote this note from a single call. This version rests on many calls. Real calls: 233 calls and 1,468 caller turns over four weeks. Load-test stand: 97,517 caller turns with simulated callers at every load level. The agent logs every evaluation of the model together with what happened next.
| Measure | Figure |
|---|---|
| Caller turns behind these figures, real calls and load test | ≈ 99,000 |
| Caller turns on real calls closed early by the model | 87% |
| Of those, closed on the model’s first look | 95% |
| Time given back per caller turn on real calls, on average | ≈ 0.3 s |
| Mid-thought pauses the model waited out | 96% |
The trade-off of a silence timer
A voice agent that only listens for silence has to choose between two bad options. If it treats any pause as the end of a turn, it interrupts people who stop to think: “My order number is… one moment… 4417.” If it waits long enough to be safe, every reply starts late, and the whole call feels sluggish. A fixed timer cannot be right for both: a short pause after “Hello?” and a short pause after “My order number is” sound the same to it.
Listening to how the sentence ends
Our agent does not decide on silence alone. After a short pause, 192 ms by default, it asks a small model a different question: from the intonation of what was just said, is the thought complete?
- If the model is confident, the turn closes at once and the reply begins. Its verdict applies on the next 32 ms frame of audio.
- If it is not, the agent keeps listening and asks again a little later.
- If the caller has stopped, the silence timer still closes the turn, so nobody is left waiting.
The model closes most turns, and the timer stays as a safety net.
On real calls
Over four weeks, 233 real calls and 1,468 caller turns:
- The model closed 87% of caller turns before the silence timer would have.
- It closed 95% of those on its first look, at the first pause it was asked about.
- When it closed a turn, it was sure: the median confidence was 0.977, and nine in ten closes were above 0.91.
- Averaged over every caller turn, including those the timer closed, the agent answered about 0.3 s sooner (292 ms).
On real calls the model closed 87% of caller turns before the silence timer, and the agent answered about 0.3 s sooner per turn.
The saving depends on how patient the agent is set to be. Each agent has two settings: the pause before the model is first asked, and the silence timer behind it. A patient agent, for callers who think aloud, waits longer, and the model then saves more because the timer it beats is longer.
| Setting: first look / timer | Caller turns | Closed early | Saved per early close |
|---|---|---|---|
| 192 / 512 ms (default) | 1,048 | 88% | 288 ms |
| 320 / 800 ms | 274 | 91% | 448 ms |
| 320 / 928 ms (most patient) | 144 | 76% | 576 ms |
It lets people think
The model also has to know when not to answer.
In 157 turns on real calls the caller paused in mid-thought and then carried on. In 151 of them, 96%, the model judged the thought unfinished and waited. Where the caller went on speaking, the model’s median confidence that they had finished was 0.015, far from a mistake.
The model tells the two situations apart by sound alone. A falling, closing intonation and a hanging, unfinished one carry different information, and the model was trained to hear that difference.
Under load
On the load-test stand, simulated callers held conversations with a recruiting agent at every load level up to 176% of a Server Unit’s rated capacity. Over 97,517 caller turns:
- the model closed 88% early;
- it closed 94% of those on the first look;
- each early close saved 0.45 s at the median, with the patient setting used on the stand (320 / 800 ms).
These figures cover every load level of the test together, from a single call to 176% of rated capacity.
Watch first, then switch on
The end-of-turn model has three modes for each agent: off, observe and commit.
In observe mode it changes nothing. It logs every judgment it would have made, together with what happened next: whether the caller stopped or carried on. That log measures how often the model would have cut someone off, on real traffic, before it is allowed to cut anyone off. The log keeps every judgment, confident or not, so we can move the threshold in either direction afterwards.
When the log agrees with what callers did, we switch the agent to commit. The figures in this note come from that log.
Where we listen hardest
- Lists and numbers. A caller reading out a card number or an address sounds finished between groups. These are the turns we sample first when we tune.
- Noisy lines. A cough or a background voice can open a turn. The agent recognizes that there is nothing to answer and stays quiet.
A companion note, Listening while talking, covers what the agent does with the caller’s channel while it speaks: early start, interruptions and short acknowledgements.