What "barge-in" is and why it makes an AI call feel human

If you've ever talked to an IVR system — "press 1 for billing, press 2 for support" — and tried to say "billing" while the list was still playing, you've experienced what happens when a system doesn't support barge-in: it ignores you and keeps talking.

Barge-in is the capability that lets a caller interrupt the AI mid-sentence, and have the AI actually stop, listen, and respond. It's a small thing that makes a large difference in how human the conversation feels.

Why it matters

In a natural conversation between two people, interruptions are normal. You cut in when you already know the answer to a question, when you want to give more context, or when the other person is heading in the wrong direction. The other person stops, adjusts, and responds to what you actually said.

Without barge-in, an AI voice agent behaves more like a recorded message than a conversation. The caller has to wait for the AI to finish its full turn before they can speak. If the AI is giving a long response — reading out a list of available appointment times, for example — the caller may be ready to answer after the first two options, but they have to sit through the rest before the system will listen to them.

This creates friction. It also signals to the caller that they're not having a real conversation, which affects their engagement with the system.

How barge-in works technically

Televox is built on LiveKit, which handles the real-time audio transport. Voice activity detection (VAD) runs continuously using the Silero VAD model — it's listening for the presence of speech at all times, even while the AI is talking.

When the VAD detects that the caller is speaking while the AI's audio is playing, it signals the system to stop the current TTS playback and end the AI's turn. The system then processes what the caller said and generates a new response.

The technical challenge is distinguishing between the caller speaking and other sounds (background noise, audio bleed from the TTS output itself). This is why VAD model quality matters — false positives (stopping mid-sentence when the caller didn't actually speak) are disruptive, and false negatives (failing to detect real speech) defeat the purpose.

Semantic turn detection

Barge-in is one side of the turn-detection problem. The other side is knowing when the caller has actually finished speaking. If the system cuts in too early — during a natural mid-sentence pause — it interrupts the caller. If it waits too long to confirm the utterance is complete, it adds unnecessary latency.

Televox uses a semantic turn-detector model that evaluates whether an utterance is likely complete based on its content, not just whether there's been a silence. This reduces the false-positive interruption rate (system cuts in during a pause) and lets the endpointing be more responsive without sacrificing accuracy.

Why this is harder than it sounds

Real phone calls have background noise: traffic, TVs, dogs, other people talking. The microphone on a mobile phone picks up ambient sound. A plumber calling from a job site has a compressor running in the background.

Getting barge-in to work reliably across these conditions — not triggering on background noise, not missing real speech — requires the VAD to be tuned carefully. Too sensitive, and background noise breaks every conversation. Not sensitive enough, and the caller has to wait through long AI monologues before they can speak.

We tune this continuously based on real call data from the observability system, which logs VAD events and turn transitions.

What it feels like when it works

When barge-in is working well, you don't notice it. The conversation flows. You say something, the AI responds. If it starts saying something you don't need to hear, you can cut in and redirect it. That's what natural conversation feels like — and that's the goal.