Barge-in
Barge-in is a caller speaking over the agent and the agent stopping. It works on Connect's realtime calls, and it fails in exactly two directions: not stopping, which callers experience as being talked over, and stopping for something that was not an interruption. An interruption still running past two seconds is recorded as ignored.
The two failure directions#
| Failure | The caller's experience | What it usually is |
|---|---|---|
| The agent does not stop | Talking to something that is not listening. Callers repeat themselves, louder, then give up | Speech onset detected too late, or a threshold that requires more words than the caller used |
| The agent stops for nothing | A sentence that dies halfway through for no reason | A cough, a door, a 'mm-hm', or the caller's own line echoing |
| The agent stops, then resumes | Being spoken over precisely when they started answering | A false-interruption resume firing without the evidence to justify it |
The third row is the compound case and the worst of the three, because it combines a broken sentence with an interruption of the caller. It is also the one that looks like a feature in a settings panel.
Measuring it rather than arguing about it#
Every interruption is timed on the wire. When the caller speaks over the agent, the worker records how long the voice carried on afterwards and posts it as an event with the rest of the call's timings. Past two seconds, the interruption is recorded as ignored — long enough that no reasonable listener would call it responsive.
That number is what turns a subjective complaint into a review finding. The post-call review lists ignored interruptions and false stops among its deterministic findings, each paired with the setting that addresses it, and separates them from the model's own opinion on delivery, which is marked as the model's.
The controls, and the ones that backfire#
- Interruption threshold, in seconds and in words
- How much speech counts as an interruption. Raising it stops false barge-ins and starts producing ignored ones; there is no setting that removes both.
- Noise cancellation
- Reduces the number of non-speech events that reach the decision at all, which is the cheapest improvement available on a noisy line.
- False-interruption resume
- Picks the sentence back up when the interruption looks spurious. Sound in principle; it depends entirely on having evidence about what the interruption was.
- The greeting
- Spoken with the caller's audio muted, so the opening is not cut in half by a hello, a hold-music fragment or a network artefact at pickup.
A related trap is the acknowledgement token: having the host say a short 'mm-hm' while the model thinks, to make silence feel shorter. Switching that on cut the caller off and stalled the model for twenty-two seconds on a measured call. The agent barged in on itself.
Does Connect use barge-in?#
Yes, on the realtime engine, in production. A caller can talk over the voice and it stops. The turn-based carrier path cannot do this in the same way, because the carrier is reading a completed line aloud rather than streaming a model's output.
Two rules constrain it from the other side. Nothing is said into the line while the caller is speaking or was heard within the last 2.5 seconds, which prevents the agent from creating the collision in the first place. And on the default model no check-in or goodbye is spoken by the host at all — the synthesised voice has the live voice's name but not its sound, so a silence is recorded and a hang-up watcher ends true abandonment without a word rather than interrupting with one.
The design goal behind all of it is unglamorous: a caller should never have to fight for the floor, and should also never be left wondering whether the sentence that stopped is coming back.
Questions#
Why did the agent keep talking when I interrupted?
Either the interruption threshold required more speech than you used — a single word can fall under it — or speech onset was detected late on a noisy line. Both are visible after the call: the interruption is timed, and past two seconds it is listed as ignored with the setting that would change it.
Can the agent finish its sentence after a false interruption?
On the model detection path there is a resume control for exactly that. On the host semantic path it is deliberately disabled, because resuming requires knowing that the interruption was spurious, and that evidence is not available at the moment the choice is made.
Would filler sounds make the pauses feel shorter?
They were tried and rolled back. A short acknowledgement spoken by the host cut into the caller's speech and stalled the model for twenty-two seconds. The productive way to shorten a pause is to shorten it — a faster model, a smaller instruction block, a tighter end-of-turn decision.