Agent speaks the remainder of its sentence as a separate turn after barge-in

Support Help
1

When a caller interrupts the agent mid-sentence, the agent stops correctly.
But instead of responding to what the caller said, or re-asking the question,
it then speaks only the leftover words of the interrupted sentence as a turn
of its own. To the caller this sounds like a broken fragment.

Here is a real inbound call:

Call ID: call_a18bc15272789d6fc8d654e2ca3
Agent version: 75
2026-09-01, inbound, 4m 31s

From that transcript, verbatim:

Agent: What’s the best number for the team to reach you by call
User: (speaks)
Agent: or text?

The intended sentence was “What’s the best number for the team to reach you
by call or text?”. The caller heard “or text?” on its own, with no context.

What I have already ruled out:

  • Not a denoising setting. Reproduced with denoising on and off, no
    difference.
  • Not something the system prompt controls. Instructing the agent to restart
    the question after an interruption does not change it.

Settings: GPT 4.1 (default), realtime transcription on “optimize for
accuracy”, no denoising, boosted keyword list in use.

My questions:

  1. Is this known, and is a fix planned?
  2. Is there a setting that makes the agent DISCARD the remainder of an
    interrupted sentence rather than speaking it on the next turn?
  3. If not, what is the recommended way to handle barge-in so the caller does
    not hear a sentence fragment?

Happy to share the recording or my account details privately with staff.

2

Hello @davidkohlani I’m checking your Call ID with the team and will get back to you as soon as I have an update.

Thank You

3

@davidkohlani Team checked the call you shared. The “or text?” fragment was not generated as a separate new LLM response. In the transcript metadata, the sentence before the interruption and the fragment after it are both part of the same streamed response (response_id: 10): the agent was saying “What’s the best number for the team to reach you by call or text?”, the caller started speaking at “Vai,” and the remaining audio fragment “or text?” was still emitted.
A few specific findings:

  • This call was running with denoising off, so the example does not point to denoising as the cause.
  • There were no call-runtime errors logged for this call.
  • The agent’s interruption sensitivity was set to 1, which is the most sensitive setting. At that level, very short caller audio/backchannels can interrupt or split an in-progress agent sentence.

The practical mitigation is to lower interruption sensitivity and re-test. That will make the agent less eager to treat short noises/backchannels as interruptions. Prompting the agent to “restart the question after interruption” will not reliably prevent this specific artifact, because the fragment is already part of audio being streamed, not a new model response.

Thank You