Barge-in is when a user cuts in while a voice agent is talking. Handling it takes two decisions: whether the sound is a real interruption or only a backchannel, background speech or side talk, and how quickly to stop when it is real. Cascades use a separate detector for this. Full-duplex models decide inside the model.
Real interruption or not
Not every sound during the agent's turn means stop. Full-Duplex-Bench v1.5 plays four kinds of overlap while the system talks: a user interruption, a listener backchannel, the user talking to someone else, and background speech. A good system stops for the first and keeps talking through the other three, which researchers call selective interruption.
All four are speech, so a cascade that stops whenever its voice activity detector hears speech stops for all four. An interruption model is a separate part that tells them apart. LiveKit's adaptive interruption handling, for example, sorts speech during the agent's turn into a real interruption, a backchannel or noise.
Try it
The timelines show one reply from each kind of system. While the AI is talking, press Interrupt to cut in, or Say "mm-hm" to backchannel.
Keeps talking. Your interruption is not heard.
Stops after a processing delay, then answers after another pause.
Stops almost at once and answers after a short gap.
Half-duplex: Keeps talking. Your interruption is not heard. Cascade: Stops after a processing delay, then answers after another pause. Full-duplex: Stops almost at once and answers after a short gap.
The cascade here stops on any speech, as a cascade with only a voice activity detector does. With an interruption model, it can keep talking through the "mm-hm". The full-duplex lane shows what the design makes possible. For how measured models behave, see What benchmarks found.
How fast to stop
Once the system decides the user means it, the next question is how fast it stops. A half-duplex system does not stop at all: it finishes its reply, and what the user said meanwhile is not heard. A cascade stops when its detector fires. Detectors such as Pipecat's, Krisp's and tyger check every 20 to 80 milliseconds. A full-duplex model hears the user in the same stream it speaks in, so the model itself decides when to stop.
Benchmarks score speed and accuracy together. TurnBench reports how many interruptions a detector catches and how long it takes. SID-Bench penalises false alarms and late responses alike. Full-Duplex-Bench v1 checks whether the system responds after an interruption, how quickly, and whether GPT-4-turbo rates the response as relevant.
What benchmarks found
MultiPartyTurnBench, published in October 2026, tests turn-taking with two to five speakers. Five full-duplex agents were good at knowing when to start speaking, and about as good as guessing at stopping only for real interruptions.
Gemini-3.8-Live stopped for 97.7% of backchannels, as if they were interruptions. MiniCPM-o-4.5 stopped for only 1.4% of real interruptions. On the same test, interruption detection either missed real interruptions or fired too often.
Self-reported numbers can look much better. Qwen3-Omni-D2, a half-duplex model converted to full-duplex with the D2 method, reports 99.5% interruption success. Scores from different tests cannot be compared, and a high turn-taking score says little about barge-in, so test interruptions on your own calls.
Interruption models
7 components in the database handle interruptions, as their main job or alongside turn detection.
| Component | What it outputs | Delivery | Licence |
|---|---|---|---|
tygerNineNineSix | Every 80 ms: listen, hold, speak or backchannel, with barge-in, backchannel and endpoint scores | Open weights | OpenMDW-1.1 |
VoxTurnUniversity of Melbourne | End of turn yes or no, plus complete, incomplete, backchannel or wait, from the last 8 seconds of audio | Open weights | CC-BY-NC-4.0 |
Echo Omni DuplexZero Runtime | A running estimate of who holds the floor, plus end-of-turn and interruption events with confidence | Hosted API | Proprietary |
| Whether speech during the agent's turn is a real interruption or a backchannel or noise | Platform option | Proprietary | |
| Every 40 ms, the probability that the user means to interrupt rather than backchannel | SDK | Proprietary | |
Pipecat backchannel detectorPipecat | Every 20 ms while the bot speaks: backchannel or taking the floor | Open weights | BSD-2-Clause |
| barge, speech_onset and speech_over events while the agent plays audio | Open weights | CC-BY-NC-4.0 |
Compare them with the turn detectors and voice activity detectors on Components. Guide: Turn detection
Benchmarks
11 full-duplex benchmarks in the database test interruptions, by tag or by a reported metric. Most recently updated first.
| Benchmark | Main metrics | Updated |
|---|---|---|
| TurnBench Sesame AI Labs / Mundo AI / CMU / oto | end-of-turn recall, interruption recall, false-positive rate | 2026-11 |
| DuplexAct Dong et al. | immediate execution success, follow-through success | 2026-10 |
| MultiPartyTurnBench Mathur and Manocha | end-of-turn recall at fixed FPR, interruption recall at fixed FPR, signed latency | 2026-10 |
| M3-DuplexBench NTT (Fukuda et al.) | turn-taking rate and latency, pause handling, backchannel and barge-in response | 2026-10 |
| ECHO Zhao et al. | pair accuracy, yield accuracy, keep accuracy | 2026-09 |
| SID-Bench Xia et al. | Average Penalty Time (APT), false alarms, response delay | 2026-03 |
| FLEXI Ge et al. | latency, response quality, emergency interruption | 2025-09 |
| Full-Duplex-Bench v1.5 NTU / UC Berkeley / UW / MIT | overlap duration, mutual-silence rate, interruption recovery | 2025-08 |
| FD-Bench Peng et al. | interruption response, delay, robustness under noise | 2025-07 |
| Talking Turns Apple / CMU | turn-event F1, backchannel rate, interruption rate | 2025-04 |
| Full-Duplex-Bench v1 NTU / UC Berkeley / UW / MIT | turn-taking latency, back-channel timing, interruption success | 2025-03 |
Guide: How full-duplex voice AI is evaluated
FAQ
- Is barge-in the same as an interruption?
- Yes. Barge-in means cutting in while a system is speaking. Interruption handling is how the system responds: whether it stops, and how quickly.
- Should a voice agent stop for every sound?
- No. A backchannel such as "mm-hm", side talk and background speech do not ask for a turn, and stopping for them breaks off the reply for nothing. Interruption models are built to tell these apart from a real interruption.
- Do full-duplex models handle interruptions better than cascades?
- Both can notice the user while they talk. In a cascade a separate detector listens, and the language model does not hear the audio. A full-duplex model hears the user in the same stream it speaks in. Measured results vary: in MultiPartyTurnBench, five full-duplex agents did about as well as guessing at stopping only for real interruptions.
- Which interruption models have open weights?
- 4 of the 7 interruption models in the database: tyger, VoxTurn, Pipecat backchannel detector, Dialt barge-in detector. Check each licence, because some are for non-commercial use only.
Updated