A backchannel is a short listener sound, such as "mm-hm", "right" or "uh-huh", made while the other person keeps talking. It shows you are following without taking the turn. A half-duplex system cannot make one, because it is silent while it listens, and in a 2026 benchmark most full-duplex models almost never gave one when it was due.
What a backchannel does
Backchannels are one of four behaviours that a walkie-talkie exchange cannot do, along with barge-in, overlap and co-completion. The listener signals attention without asking for the floor. A system that cannot do this stays silent while you talk, and the silence reads as indifference.
Listeners use many short forms. The Backchannel Cue model chooses among 61 of them, from "mhm" to "oh no".
Where people place them
The listener says "mm-hm" while the speaker keeps the turn, at moments that leave room for it.
Short backchannels such as "hai", "un" and "sou desu" come roughly every one to two seconds, more often than in English.
In DuplexAct, most full-duplex models almost never gave a backchannel at the moment one was due.
Full-Duplex-Bench v1 compares when a model backchannels with when backchannels occur in a reference corpus, ICC, and scores the gap between the two timing distributions with Jensen-Shannon divergence. The smaller the gap, the better.
How often people backchannel depends on the language. In Japanese, short backchannels such as "hai", "un" and "sou desu" come roughly every one to two seconds, much more often than in English. A test tuned to English norms can count correct Japanese behaviour as errors.
Turn-detection models treat a backchannel as its own state. Easy Turn and VoxTurn label each moment as complete, incomplete, backchannel or wait.
How full-duplex models score
DuplexAct, published in October 2026, scored 14 models on eight interaction actions, using 4,691 test items, 2,531 of them from real recordings. Each action was scored at the moment it was due. For backchannels, success was near zero for most models.
Other entry actions gave mixed results. When the right move was to interrupt, Moshi succeeded 67.5% of the time, GPT-Realtime-2.1 8.8% and Grok Voice 0.0%. The four closed models entered the floor correctly 77.7 to 97.0% of the time on the Recover action but only 27.5 to 41.3% on Take Turn, so skill at one entry action did not predict skill at another.
A week earlier, Coupled but Late found a related timing gap. When two copies of PersonaPlex talked to each other, the floor changed hands at a median of 400 to 560 ms, against 137 ms for people in the Switchboard telephone corpus. The authors read this as the model waiting for the end of a turn instead of predicting it. Source: Signals 2026-W42.
Making them and hearing them
A voice agent meets backchannels in two directions. As a listener, it should give them while the user talks. As a speaker, it should hear the user's "mm-hm" and keep going. In MultiPartyTurnBench, Gemini-3.8-Live stopped for 97.7% of backchannels, treating them as interruptions. Interruption and barge-in covers that side.
Training data shapes both. On a single mixed recording, backchannels get absorbed into the main speaker's voice, and in synthetic dialogue they appear only where the generator's rules put them. Training data explains why each speaker needs their own channel.
Backchannel models
4 components in the database are built mainly for backchannels. Some predict when the system should give one. Others tell whether the user's speech is a backchannel or a bid for the turn.
| Component | What it outputs | Delivery | Licence |
|---|---|---|---|
Easy TurnASLP Lab, Northwestern Polytechnical University | One of complete, incomplete, backchannel or wait, from the user's audio | Open weights | Apache-2.0 |
MaAIKyoto University | Near-future voice activity for user and system, plus backchannel and nodding predictions | Open weights | Custom terms |
Pipecat backchannel detectorPipecat | Every 20 ms while the bot speaks: backchannel or taking the floor | Open weights | BSD-2-Clause |
Backchannel Cue 32MKolja Beigel | Whether a listener would respond now, and with what: one of 61 short forms such as "mhm" or "oh no" | Open weights | MIT |
6 more list a backchannel among their outputs, next to turn ends or interruptions: tyger, VoxTurn, Echo Omni, LiveKit adaptive interruption handling, Krisp VIVA Interruption Prediction v1 and X2-Turn-4B.
Benchmarks
6 full-duplex benchmarks in the database test backchannels, by tag or by a reported metric. Most recently updated first.
| Benchmark | Main metrics | Updated |
|---|---|---|
| DuplexAct Dong et al. | immediate execution success, follow-through success | 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 |
| 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 |
Read more: The full-duplex threshold, Why STS needs new benchmarks
FAQ
- Is a backchannel the same as an interruption?
- No. A backchannel signals attention and leaves the turn with the speaker. An interruption takes the turn. A voice agent that treats every "mm-hm" as an interruption stops talking for no reason.
- Why can't a cascade backchannel?
- Its reply starts only after the user's turn ends, so nothing in the pipeline speaks while the user is talking. A separate backchannel model can add them.
- Do backchannels differ by language?
- Yes. In Japanese, short backchannels such as "hai", "un" and "sou desu" come roughly every one to two seconds, much more often than in English. A test tuned to English norms can count correct Japanese behaviour as errors.
- Which benchmarks measure backchannels?
- 6 full-duplex benchmarks in the database do, including Full-Duplex-Bench v1, which compares backchannel timing with human recordings, and DuplexAct, which checks whether a model gives one when it is due.
Updated