Skip to main content
All articles
Published on6 min read

When the caller cuts off the AI voice agent: the barge-in handling that separates good agents from bad ones

A patient caller listens. A rushed caller interrupts. If the agent keeps talking over them, you lose the call. Here's barge-in handling in 3 parameters.

'My AI voice agent talks over me.' That's the first complaint that makes a customer hang up and never call back. The technical term is 'barge-in': the agent's ability to stop immediately when the caller starts speaking. Badly tuned, it's the #1 experience killer. Here are the 3 parameters that fix it.

Parameter 1 — voice detection threshold#

The agent listens in parallel to its own speech. Too sensitive: it stops on background noise (café, car, breathing), restarts, sounds unstable. Not sensitive enough: it keeps talking 1.5s after the caller has begun. Sweet spot: trigger at 150-200ms of continuous detected voice, sound energy > 25dB above background.

Parameter 2 — cut-off duration#

Once barge-in is detected, the agent must stop within 80-120ms. Past 200ms the user perceives an overlap that frustrates them. Below 50ms the agent stops too often (false positives on 'hmm'). Measure with a simple test: 10 calls where you interrupt at the 3-second mark. Target: zero perceptible overlap.

Parameter 3 — contextual resume#

Once interrupted, the agent listens to the new content and does NOT resume its original sentence. Bad: 'As I was saying, your appointment...'. Good: hears the new info, integrates it, answers to that only. This requires a context-aware LLM with multi-turn state, not a simple FIFO queue of pre-generated phrases.

The 4 barge-in cases the agent must handle#

  • Question during explanation: interrupt cleanly, answer the question, offer to resume.
  • Correction of misheard info: confirm, re-engage.
  • Early answer ('yes that's it'): skip the planned confirmation, go to next step.
  • Urgent request ('wait, it's urgent'): abandon current flow, emergency mode + handoff.

The quick 5-call test#

Call your own agent and cut it off 5 times at different moments: start of sentence, middle, end. If every time it stops in < 200ms AND resumes by answering YOUR input (not repeating its own), your barge-in is set. Otherwise, tune the parameter before production. First month VocazAI free to calibrate.