Turn-taking is how speakers pass the turn: when to start, when to wait, when to give way and when to say "mm-hm". A full-duplex model decides this continuously, and recent benchmarks show where it still differs from people. This page collects the numbers, benchmarks, corpora and open models for studying it, with a source for each.
Human conversation in numbers
- Across ten languages, responses to yes/no questions began about 200 ms after the question ended on average (mean 208 ms). In every language the most common offset was between 0 and 200 ms (Stivers and colleagues, 2009).
- In three conversational corpora, in Dutch, Scottish English and Swedish, the most common interval at a speaker change is a gap of about 200 ms, and about 40 percent of the intervals are overlaps (Heldner and Edlund, 2010).
- Producing even one word takes about 600 ms from planning to speech, a figure from picture-naming studies. Levinson and Torreira (2015) argue that listeners must therefore plan a reply and predict the end of the turn while still listening.
- On the telephone calls of the Switchboard corpus, the median time for the turn to pass from one person to the other is 137 ms, by the rule used in Coupled but Late (Zhu and colleagues, 2026). This is a median, so it is a different measure from the most common gap above.
- Listener responses such as short comments and head movements are far more frequent in Japanese casual conversation than in American English (Maynard, 1990), so backchannel rates measured on English do not carry over as they are.
Read more: The full-duplex threshold, Why STS needs new benchmarks
Open problems, November 2026
- One open model hands over late. In Coupled but Late (Zhu and colleagues, 2026), two copies of PersonaPlex talking to each other passed the turn after a median of 400 to 560 ms, against a median of 137 ms between people on Switchboard. When the authors delayed one direction of the audio, the reply moved by the same amount, which suggests the model reacts to the end of a turn where people predict it. The study covers one model.
- Models rarely give backchannels. DuplexAct (Dong and colleagues, 2026) scored 14 models on eight interaction actions, each at the moment it was due. For backchannels, Raon-SpeechChat, which has a dedicated backchannel token, led at 31.0%, and GPT-Realtime-2.1 scored 0.2%. Talking Turns reported in 2025 that speech models rarely give backchannels.
- Models do not tell interruptions from other overlapping speech. In MultiPartyTurnBench (Mathur and Manocha, 2026), on 240 synthetic conversations, five full-duplex agents scored 86.1 to 95.8% balanced accuracy on when to start, and 46.9 to 54.7%, close to guessing, on stopping only for real interruptions. Gemini-3.8-Live stopped for 97.7% of backchannels, and MiniCPM-o-4.5 stopped for 1.4% of real interruptions.
- Models struggle when the user corrects themselves. In the Full-Duplex-Bench v3 paper (April 2026), GPT-Realtime did best of six systems and passed 58.8% of the scenarios in which the user corrected a request mid-sentence. A cascade passed 17.6%.
- No public benchmark scores a model supplying the word a speaker is searching for (co-completion). Duplex Cue (September 2026) scores the other direction: whether a model adapts when the listener supplies a missing word, a correction or a clarification. On 66 cues, people adapted 68.2% of the time and PersonaPlex 34.8%.
Guides: Turn detection, Interruption and barge-in, Backchannels
Benchmarks
24 full-duplex benchmarks in the database name a turn-taking behaviour in their tags or listed metrics, most recently updated first. The Tests column is built from those tags and metric names, so it can miss a behaviour that a paper tests and the entry does not list. Each benchmark page has the systems tested and the results.
| Benchmark | Tests | Paper | Updated |
|---|---|---|---|
| TurnBench Sesame AI Labs / Mundo AI / CMU / oto | turn-taking, turn ends, interruptions | arXiv | 2026-11 |
| Duplex-MPE Ma et al. | several speakers, who is addressed | arXiv | 2026-11 |
| eot-bench LiveKit | turn-taking, turn ends | — | 2026-11 |
| Cekura Turn-Taking Benchmark Cekura | turn-taking, turn ends, interruptions, backchannels, pauses | — | 2026-11 |
| DuplexAct Dong et al. | turn-taking, interruptions, backchannels | arXiv | 2026-10 |
| MultiPartyTurnBench Mathur and Manocha | turn-taking, turn ends, interruptions, several speakers | arXiv | 2026-10 |
| M3-DuplexBench NTT (Fukuda et al.) | turn-taking, interruptions, backchannels, pauses | arXiv | 2026-10 |
| MP-Bench Shih et al. | turn-taking, several speakers | arXiv | 2026-10 |
| MultiTalkBench Wang et al. (MultiTalk) | turn-taking, several speakers, who is addressed | arXiv | 2026-09 |
| ECHO Zhao et al. | turn-taking, interruptions, backchannels | arXiv | 2026-09 |
| DuplexSpeechBench-IFEval Mathur and Manocha | turn-taking rules set by instructions or persona | arXiv | 2026-09 |
| DuplexWorld Centific / University of Maryland | turn-taking | arXiv | 2026-08 |
| Game-Time Benchmark MIT / NTU / Academia Sinica | turn-taking, overlap | arXiv | 2026-06 |
| Full-Duplex-Bench v3 NTU (NVIDIA advisory) | turn-taking, disfluent speech | arXiv | 2026-05 |
| OmniInteract CUHK MMLab / SJTU / NTU / McMaster / CityUHK / JUFE | interruptions | arXiv | 2026-05 |
| CoDeTT Bairong (BRVoice team) | turn-taking, interruptions, backchannels | arXiv | 2026-04 |
| SID-Bench Xia et al. | interruptions, backchannels | arXiv | 2026-03 |
| Full-Duplex-Bench v2 NTU / UC Berkeley / UW / MIT | overlap | arXiv | 2026-02 |
| HumDial (ICASSP '26) ICASSP 2026 Spoken Dialogue Challenge | turn-taking | arXiv | 2026-01 |
| FLEXI Ge et al. | interruptions | arXiv | 2025-09 |
| Full-Duplex-Bench v1.5 NTU / UC Berkeley / UW / MIT | turn-taking, interruptions, backchannels, overlap | arXiv | 2025-08 |
| FD-Bench Peng et al. | interruptions | arXiv | 2025-07 |
| Talking Turns Apple / CMU | turn-taking, turn ends, interruptions, backchannels | arXiv | 2025-04 |
| Full-Duplex-Bench v1 NTU / UC Berkeley / UW / MIT | turn-taking, interruptions, backchannels, pauses | arXiv | 2025-03 |
A chart of which benchmark covers which behaviour is in How full-duplex voice AI is evaluated. Scores by benchmark version and by who ran it: Analytics.
Two-channel corpora
Overlap and backchannels are measured most reliably, and learned by today's full-duplex models, from recordings with each speaker on their own channel. 40 datasets in the database keep each speaker on their own channel, and 32 of them publish their size. Of those 32, 2 are synthetic speech, and DuplexChat was split into two channels from podcast audio after recording.
18 can be used for research at no cost: 10 under terms that also allow commercial use and 8 under non-commercial terms. Both lists are in How to choose a full-duplex training dataset. The largest free recordings of human speech are Seamless Interaction (4,000+ h), SmoothConv / DuplexConv (2,100 h), Open Yap 1K (1,000 h), CANDOR (~860 h) and otoSpeech-full-duplex-280h (280 h).
The database tags 8 of the 40 for turn-taking, backchannels or turn ends: Switchboard-1, CANDOR, CALLHOME, CALLFRIEND, HKUST Mandarin Telephone Speech, Part 1, otoSpeech-full-duplex-turn-104h, Real-TurnTurk, SmoothConv / DuplexConv.
Guide: Training data for full-duplex models
Models and components
- 27 full-duplex models have public weights, so their turn-taking can be measured and changed. They are listed with licences under Open models you can run.
- Turn-taking is also built as separate parts for cascaded agents. The database lists 24 turn detectors, 9 voice activity detectors, 4 interruption models and 2 backchannel models under Components, counted by main job. 3 more also handle backchannels, and 5 more also handle interruptions.
- Voice Activity Projection projects the voice activity of both speakers over the next two seconds, and shift, hold and backchannel decisions are derived from that.
- Two shared tests rank parts like these. TurnBench scores turn-end and interruption detection on 30 hours of studio-recorded two-person conversation, labelled by hand. LiveKit's eot-bench scores how long a detector waits after the user stops. Results from both are in Turn detection.
Keeping up
- The weekly report lists new papers, models and datasets on full-duplex speech.
- The whole database is a Hugging Face dataset under CC BY 4.0.
FAQ
- What is turn-taking in voice AI?
- Turn-taking is how speakers pass the turn to each other: when one starts, stops, pauses, overlaps or acknowledges. A full-duplex model has to make that decision continuously, because it listens while it speaks.
- Which benchmark should I use for turn-taking?
- It depends on the behaviour. Full-Duplex-Bench v1 and v1.5 test timing, interruptions and backchannels between two speakers on fixed recordings. TurnBench scores turn-end and interruption detection on their own. MultiPartyTurnBench covers two to five speakers, and DuplexAct scores each action at the moment it is due. 24 full-duplex benchmarks in the database name a turn-taking behaviour in their tags or listed metrics.
- Which datasets have two channels for turn-taking research?
- 40 datasets in the database keep each speaker on their own channel, and 32 of them publish their size. 18 can be used for research at no cost: 10 under terms that also allow commercial use and 8 under non-commercial terms. The largest free recordings of human speech are Seamless Interaction (4,000+ h), SmoothConv / DuplexConv (2,100 h), Open Yap 1K (1,000 h), CANDOR (~860 h) and otoSpeech-full-duplex-280h (280 h).
- Is there turn-taking research on Japanese?
- Yes. Listener responses are far more frequent in Japanese casual conversation than in American English (Maynard, 1990), so a test tuned to English can count correct Japanese behaviour as an error. M3-DuplexBench is the one full-duplex benchmark in the database that scores Japanese turn-taking. It scores turn-taking, pauses, user backchannels and barge-in, and its listed measures do not include the model's own backchannels. Two-channel datasets in the database that include Japanese: CALLHOME, CALLFRIEND, Japanese Duplex Conversation Dataset, DuplexChat, Global Conversational Speech, Basis Conversations 1500 and DAI-ASR-I18N. J-Moshi and LLM-jp-Moshi-v1 are open Japanese models to test.
Updated