Why 120 Milliseconds Matters: The Science of Barge-In

What barge-in actually is

Barge-in is the moment when the caller starts speaking while the agent is still talking. Either the caller wants to interrupt, change direction, or correct the agent. A natural conversation handles this without anyone having to wait for a slot to open. The agent that handles barge-in well stops, listens, and adjusts.

The 120-millisecond rule

Human reaction time to a stimulus is roughly 150 to 200 milliseconds. If a voice agent takes longer than that to detect and respond to interruption, the conversation feels off. The caller feels they are talking over a machine that is not listening, not a partner that is responsive.

Detecting the interruption

The agent listens to incoming audio continuously, not just during its own speech pauses. The detection has to distinguish between the caller actually speaking and ambient sounds, background voices, or environmental noise (a car passing, a baby crying, a television in the background).

Stopping mid-word

When the agent detects an interruption, it cannot finish its current sentence. It has to stop within milliseconds, mid-syllable if necessary. The audio output flushes and the agent is ready to listen.

Resuming with context

After the caller input is processed, the agent has to figure out whether to continue what it was saying, abandon it, or pick up a new thread. This is a reasoning step, not a script.

What Edah AI does

Edah AI handles barge-in in under 120 milliseconds end-to-end, from the first frame of caller audio to the stop of agent output. The numbers are measured on every release and tracked across percentiles, not just averages. A good 99th-percentile barge-in is what makes the conversation feel like it has two participants, not one performer.