In 30 hours of hand-labeled two-person conversation, listeners began speaking a median 151 ms before the current turn had ended, while Amazon Lex V2, by default, waits 600 ms of silence after a caller stops before it decides the utterance is complete. That gap is the whole problem of end-of-turn detection: people predict the end of a turn, most machines wait for silence. Waiting has a measurable cost in both directions. A robot using silence-based end-of-turn detection interrupted its users in 16.6% of its turns and still answered with a median delay of 2.7 s, and when users talked over commercial voice models, the time those models needed to stop speaking ranged from 0.23 s to 2.25 s. This roundup focuses on the turn decision itself; for the full pipeline delay from microphone to first audio, see our voice AI agent latency statistics. We aggregated data from the TurnBench benchmark, the Max Planck PNAS turn-taking study, Amazon Science and Google Research endpointing papers, Kyutai’s Moshi paper, the Full-Duplex-Bench and Talking Turns benchmarks, and peer-reviewed Voice Activity Projection research.
TL;DR
- Listeners start speaking a median 151 ms before the other person’s turn ends in smooth handovers (TurnBench, IEEE SLT 2026)
- 51% to 55% of turn transitions across three conversation corpora take less than 200 ms (Heldner and Edlund 2010, reviewed in Frontiers in Psychology 2015)
- Amazon Lex V2 waits a default 600 ms of silence before treating a spoken utterance as finished (AWS Lex V2 documentation)
- Joint speech recognition and endpointing cut Google’s median endpoint latency from 390 ms to 270 ms (Bijwadia et al., Google, SLT 2022)
- A silence-based robot baseline interrupted users in 16.6% of system turns versus 6.9% for a predictive model (Skantze and Irfan, KTH, 2025)
- A turn-taking model trained on English fell to 59.43% balanced accuracy on Japanese conversation (Inoue et al., LREC-COLING 2024)
- When users interrupted, GPT-4o Realtime stopped talking in 0.23 s and Nova Sonic in 2.25 s (Full-Duplex-Bench v1.5, 2025)
- Only 17.4% of user interruptions ended with the user taking the floor from Moshi, versus 63.6% in human conversation (Arora et al., ICLR 2025)
- Moshi’s full-duplex design has a theoretical latency of 160 ms and 200 ms in practice (Kyutai, Moshi paper 2024)
- A cascaded dialogue pipeline produced 0.0 s of overlapping speech per minute, against 3.6 s in real Fisher phone calls (Nguyen et al., Generative Spoken Dialogue Language Modeling 2022)
1. The Human Benchmark: Turns Overlap More Than They Wait
The standard picture of conversation, one person talks, stops, the other answers, is wrong in a way that matters for software. In TurnBench, a 2026 benchmark built from 30 hours of hand-labeled dyadic (two-person) conversation, listeners began speaking a median 151 ms before the current turn ended, and once floor-taking interruptions are included the median offset reaches 281 ms before the end (TurnBench, Jiang et al., IEEE SLT 2026). A “floor transfer” here simply means the moment the right to speak passes from one person to the other. A detector that waits for silence is late by design, before any processing even starts.
Timing also carries meaning. The PNAS 10-language study found that responses that do not answer the question (for example “I don’t know”) arrive 131.78 ms later than real answers, and confirmations arrive 206.87 ms sooner than disconfirmations in the multivariate model (Stivers et al., PNAS 2009). Humans hear those delays as signals. A voice agent that always waits the same fixed silence before answering sends a signal it does not intend.
| Metric | Value | Source |
|---|---|---|
| Median offset when the listener starts speaking, smooth handovers (negative = before the turn ends) | -151 ms | TurnBench, Jiang et al., IEEE SLT 2026 |
| Most common (modal) response offset after a yes/no question, 10 languages combined | 0 ms (per-language modes 0 to +200 ms) | Stivers et al., PNAS 2009 |
| Extra delay when the response does not answer the question, multivariate model | +131.78 ms (95% CI 59.34 to 204.23) | Stivers et al., PNAS 2009 |
| Delay difference for confirmations vs disconfirmations, multivariate model | -206.87 ms (95% CI -268.61 to -145.12) | Stivers et al., PNAS 2009 |
| Mean response offset among responses that coders judged to be on time, Japanese vs Danish | 36 ms vs 203 ms | Stivers et al., PNAS 2009 |
| Share of turn transitions faster than 200 ms, three corpora | 51% to 55% | Heldner and Edlund 2010, reviewed in Levinson and Torreira, Frontiers in Psychology 2015 |
| Share of the Switchboard signal where both speakers talk at once | 3.8% | Levinson and Torreira, Frontiers in Psychology 2015 |
Context note: the overlap picture depends on how it is measured. Overlapping speech covers only 3.8% of the Switchboard audio, yet Heldner and Edlund found overlaps in about 40% of transitions, because most overlaps are short (Levinson and Torreira, Frontiers in Psychology 2015). Most recent available data for the cross-language figures: Stivers et al., 2009, and Heldner and Edlund, 2010. Human turn timing changes slowly, so the age of these studies matters less than for product benchmarks.
2. How Long Machines Wait: Silence Timeouts and Endpoint Latency
Most production systems still decide that a turn has ended with a timer. Amazon Lex V2 waits a default 600 ms of silence after the caller stops speaking before it treats the utterance as finished, waits 4,000 ms for the caller to start at all, and truncates any single utterance at 55,000 ms (AWS, Configuring timeouts for capturing user input with a Lex V2 bot). Research papers describe even longer settings: the Voice Activity Projection authors say spoken dialogue systems usually use a silence timeout of around 1 or 2 seconds.
The research frontier is shaving milliseconds off the moment the system commits. Google’s joint model, which shares layers between speech recognition and the endpointer (the component that decides the user is done), cut median endpoint latency from 390 ms to 270 ms and 90th-percentile latency from 740 ms to 570 ms, with word error rate held at 5.8% (Bijwadia et al., Google, Unified End-to-End Speech Recognition and Endpointing, SLT 2022). Faster commits only help if the transcript is right, a theme covered in our speech recognition hallucination statistics.
| Metric | Value | Source |
|---|---|---|
| Default silence wait after the caller stops speaking, Amazon Lex V2 | 600 ms | AWS Lex V2 documentation |
| Default silence delay before a VAD-only turn end, LiveKit Agents (vendor default) | 500 ms | LiveKit blog, December 20, 2024 (vendor’s own setting) |
| Usual silence timeout in spoken dialogue systems | Around 1 or 2 seconds | Inoue et al., Kyoto University, IWSDS 2024 |
| Median / 90th-percentile endpoint latency, separate recognition and endpointing models (Google voice search, about 14,000 utterances) | 390 ms / 740 ms | Bijwadia et al., Google, SLT 2022 |
| Median / 90th-percentile endpoint latency, joint model | 270 ms / 570 ms (120 ms and 170 ms lower, 30.8% and 23.0%) | Bijwadia et al., Google, SLT 2022 |
| Endpoint latency at the 50th / 90th / 99th percentile, Alexa baseline on transactional queries | 300 / 570 / 1,110 ms | Raju et al., Amazon, Two-Pass Endpoint Detection 2024 |
Context note: the percentile spread is the story. A median of 300 ms looks close to human timing, but the slowest 1% of Alexa transactional queries waited 1,110 ms or more for the endpoint decision alone, before any answer is generated (Raju et al., Amazon, Two-Pass Endpoint Detection for Speech Recognition, 2024). Most recent available data for the Google figures: 2022.
3. Cut-Ins vs Late Replies: The Endpointing Trade-Off
Every endpointer trades two failures against each other: cutting the user off mid-thought (a false cut-in) or leaving a dead silence after the user is done. The most direct measurement we found comes from a 39-participant study with a Furhat robot head. With the standard silence-based baseline, the robot interrupted users in 16.6% of its turns and still had a median response time of 2.7 s; a predictive model combining TurnGPT and Voice Activity Projection cut interruptions to 6.9% and the median response to 1.5 s, both differences at p < 0.001 (Skantze and Irfan, KTH, Applying General Turn-taking Models to Conversational Human-Robot Interaction, 2025). The silence-based design lost on both axes at once.
Fixing one side usually costs the other. Amazon’s second-pass “arbitrator”, which double-checks the first endpointer before it closes the microphone, cut early endpointing (stopping before the user had finished) by 16.34% on transactional queries, but pushed 99th-percentile latency from 1,110 ms to 1,320 ms. The bluntest detectors fail in the other direction: simple sound-level triggers fired on more than half of the moments when the turn was not actually over (see the false positive row below). Those triggers react to sound rather than meaning, so the audio that reaches them matters, which connects to our neural noise suppression statistics.
| Metric | Value | Source |
|---|---|---|
| System turns that interrupted the user, silence-based baseline vs TurnGPT plus VAP model | 16.6% vs 6.9% | Skantze and Irfan, KTH, 2025 |
| Median system response time, silence-based baseline vs TurnGPT plus VAP model | 2.7 s vs 1.5 s | Skantze and Irfan, KTH, 2025 |
| Participants who preferred the silence-based baseline / had no preference | 9 of 39 / 3 of 39 | Skantze and Irfan, KTH, 2025 |
| Utterances where the endpointer stopped listening too early, SLURP benchmark, baseline vs with arbitrator | 2.39% vs 2.19% (8.37% relative reduction) | Raju et al., Amazon, 2024 |
| Relative cut in early endpointing on transactional queries / 99th-percentile latency before vs after | 16.34% / 1,110 ms vs 1,320 ms | Raju et al., Amazon, 2024 |
| Precision when predicting a true end of turn, acoustic detector vs end-to-end model (recall 97.3% vs 97.5%) | 67.3% vs 84.7% | Chang et al., Google, Interspeech 2022 |
| False positive rate on TurnBench, simple energy detector vs OpenAI Realtime server VAD | 0.632 vs 0.525 | TurnBench leaderboard, viewed October 31, 2026 |
| Reduction in unintentional interruptions, transformer turn detector plus VAD vs VAD alone | 85% | LiveKit blog, December 20, 2024 (vendor’s own claim, no test set disclosed) |
Context note: the robot baseline relied on an undocumented Google end-of-speech detector that the authors infer is threshold-based, and they cite a 1 s threshold, so read it as a silence-based system rather than a precise timer setting. Google’s end-of-turn study used disfluent test speech (random pauses, filled pauses such as “um”, lengthened words and repeated phrases), the cases where the paper says a conventional endpointer cuts the user off (Chang et al., Google, Turn-Taking Prediction for Natural Conversational Speech, Interspeech 2022). Its natural-conversation test set has only 200 utterances from 10 speakers, so treat the precision figures as directional.
4. Predicting the End of a Turn: VAP, TurnGPT and the Leaderboard
Predictive models listen to how the sentence is going, not only to whether the room went quiet. TurnGPT, a text model, reached 0.823 balanced accuracy on spontaneous spoken dialogue against 0.748 for an LSTM baseline (Ekstedt and Skantze, TurnGPT, Findings of EMNLP 2020). Balanced accuracy averages performance on “turn shifts” and “holds”, so 0.5 is chance. Voice Activity Projection (VAP) works on the audio itself and predicts the next 2 seconds of who will speak.
Language is the hidden variable. A VAP model trained on English fell to 59.43% balanced accuracy on Japanese, against 74.20% for a Japanese-trained model, and turn shifts made up 42.9% of shift-or-hold moments in the Japanese data versus 9.9% in English (Inoue et al., Multilingual Turn-taking Prediction Using Voice Activity Projection, LREC-COLING 2024). A turn detector tuned on English calls is not a neutral component in a multilingual product.
| Metric | Value | Source |
|---|---|---|
| Balanced accuracy predicting turn shifts in spontaneous spoken dialogue, TurnGPT vs LSTM baseline | 0.823 vs 0.748 | Ekstedt and Skantze, KTH, Findings of EMNLP 2020 |
| VAP shift-vs-hold balanced accuracy, trained and tested on the same language: English / Mandarin / Japanese | 79.59% / 84.49% / 74.20% | Inoue et al., Kyoto University, LREC-COLING 2024 |
| VAP trained on English, tested on Japanese | 59.43% | Inoue et al., Kyoto University, LREC-COLING 2024 |
| Share of shift-or-hold events that are turn shifts, English vs Japanese test data | 9.9% vs 42.9% | Inoue et al., Kyoto University, LREC-COLING 2024 |
| VAP on a CPU, 1 s vs 20 s of audio context, Japanese data: real-time factor (below 1 = faster than real time) | 0.73 vs 13.69 | Inoue et al., Kyoto University, IWSDS 2024 |
| Voice Activity Projection on TurnBench: recall / false positive rate / latency | 0.845 / 0.055 / 368 ms | TurnBench leaderboard, viewed October 31, 2026 |
| Smart Turn v3 vs Kyutai Semantic VAD on TurnBench: recall / false positive rate / latency | 0.752 / 0.047 / 1,017 ms vs 0.773 / 0.059 / 1,007 ms | TurnBench leaderboard, viewed October 31, 2026 |
| Gemini 3.1 Live vs OpenAI Realtime semantic VAD on TurnBench: recall / false positive rate / latency | 0.657 / 0.022 / 1,234 ms vs 0.303 / 0.018 / 793 ms | TurnBench leaderboard, viewed October 31, 2026 |
Context note: the shift shares come from different corpora (Switchboard phone calls vs a Japanese travel-agency task), so the gap reflects task as well as language. The real-time result matters for deployment: with 1 second of context, VAP ran at 14.61 ms per frame on an Intel Xeon CPU and reached 76.16% balanced accuracy, slightly above the 74.20% measured with 20 seconds of context (Inoue et al., Real-time and Continuous Turn-taking Prediction Using Voice Activity Projection, IWSDS 2024). The TurnBench leaderboard disqualifies systems with a false positive rate above 0.15, and the two most cautious systems in the table keep false alarms lowest while missing the most real turn ends. Most recent available data for TurnGPT: 2020.
5. Barge-In: What Happens When the User Talks Over the Agent
Barge-in is the reverse problem: the agent is talking and the user starts speaking. Stopping fast is only half the job. On Full-Duplex-Bench v1.5, GPT-4o Realtime stopped speaking 0.23 s after a user interruption and addressed what the user said 78% of the time, while Nova Sonic took 2.25 s to stop and ignored the interruption to continue its prior answer 71% of the time (Lin et al., Full-Duplex-Bench v1.5, arXiv 2025). Contact-center teams evaluating these models, discussed in our voice agent contact center automation statistics, are choosing between two very different behaviors.
The opposite error is yielding to a backchannel, the short “mm-hm” or “right” a listener says to show they are following. GPT-4o Realtime stopped in 0.21 s even on those cues, which the authors read as over-eager halting. Instant yielding is not the human norm either: an interrupted person keeps talking for a median 1.48 s before ceding the floor, so the lowest stop time is not automatically the right target. Moshi sits at the far end of that scale, with only 17.4% of user interruptions ending in the user taking the floor.
| Metric | Value | Source |
|---|---|---|
| User interrupts the agent, GPT-4o Realtime: rate of addressing what the user said / rate of continuing its prior answer / time to stop speaking | 0.78 / 0.10 / 0.23 s | Lin et al., Full-Duplex-Bench v1.5, 2025 |
| User interrupts the agent, Gemini 2.0 Flash Live endpoint: rate of addressing what the user said / rate of continuing its prior answer / time to stop speaking | 0.33 / 0.55 / 2.20 s | Lin et al., Full-Duplex-Bench v1.5, 2025 |
| User interrupts the agent, Nova Sonic: rate of addressing what the user said / rate of continuing its prior answer / time to stop speaking | 0.24 / 0.71 / 2.25 s | Lin et al., Full-Duplex-Bench v1.5, 2025 |
| User only backchannels (“mm-hm”): GPT-4o Realtime time to stop vs Nova Sonic rate of continuing its prior answer | 0.21 s vs 0.98 | Lin et al., Full-Duplex-Bench v1.5, 2025 |
| User interruptions that end with the user taking the floor, by who was interrupted: Moshi / cascaded system / human speaker | 17.4% / 59.4% / 63.6% | Arora et al., Talking Turns, ICLR 2025 |
| Moshi decisions that were false-positive interruptions of the user | 2.09% of all decisions | Arora et al., Talking Turns, ICLR 2025 |
| Median time an interrupted human keeps talking before ceding the floor | 1.48 s | TurnBench, Jiang et al., IEEE SLT 2026 |
| Barge-in (allow-interrupt) default for Amazon Lex V2 prompts | On (true) | AWS Lex V2 documentation |
Context note: the Talking Turns user study collected about 4 hours of human-AI conversation per system with 11 participants, and user interruptions made up only 0.2% (Moshi) and 0.1% (cascaded) of the instances scored, so these percentages rest on small counts, the cascaded one especially (Arora et al., Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics, ICLR 2025). Full-Duplex-Bench v1.5 uses 200 synthesized interruptions, not live callers. We found no public primary dataset on how often contact-center callers barge in on production voice agents.
6. Full-Duplex Models: Listening and Speaking at the Same Time
Full-duplex models drop the turn decision entirely: they listen and speak on parallel audio streams, so overlap and backchannels can happen naturally. Kyutai’s Moshi reports a theoretical latency of 160 ms and 200 ms in practice, and its authors note that overlapping speech makes up 10% to 20% of spoken time in real conversation (Défossez et al., Kyutai, Moshi: a speech-text foundation model for real-time dialogue, 2024). The cascaded baseline in the dGSLM evaluation produced none of it, with zero pauses and zero overlap per minute of generated dialogue (Nguyen et al., Generative Spoken Dialogue Language Modeling, 2022).
Speed has not yet become judgment. On Full-Duplex-Bench, Moshi scored a takeover rate of 0.980 during user pauses in the Candor set, against 0.310 for Gemini Live; for this measure lower is better, because taking the floor while the user is only pausing is a cut-in (Lin et al., Full-Duplex-Bench, ASRU 2025). Full-duplex models fixed the silence problem and inherited the interruption problem.
| Metric | Value | Source |
|---|---|---|
| Moshi theoretical / practical response latency | 160 ms / 200 ms | Défossez et al., Kyutai, 2024 |
| Moshi audio pretraining / Fisher phone-conversation fine-tuning data | 7 million hours / 2,000 hours | Défossez et al., Kyutai, 2024 |
| Overlapping speech in real conversation, as cited in the Moshi paper | 10% to 20% of spoken time | Défossez et al., Kyutai, 2024 (citing Cetin and Shriberg 2006) |
| Overlap per minute, real Fisher conversations vs cascaded dialogue system | 3.6 s vs 0.0 s | Nguyen et al., dGSLM, 2022 |
| Between-speaker gaps per minute, real Fisher conversations vs cascaded system | 7.5 vs 14.9 | Nguyen et al., dGSLM, 2022 |
| Cumulative gap / overlap, Moshi at sampling temperature 1.0 vs real conversations (same dGSLM protocol) | 4.5 s / 4.1 s vs 4.2 s / 3.3 s | Défossez et al., Kyutai, 2024 |
| Share of conversation time spent in silent gaps, Moshi vs cascaded system (user study) | 11.8% vs 32.4% | Arora et al., Talking Turns, ICLR 2025 |
| Takeover rate during user pauses, Candor set (lower is better): Moshi / dGSLM / Gemini Live | 0.980 / 0.935 / 0.310 | Lin et al., Full-Duplex-Bench, 2025 |
Context note: Moshi’s turn statistics swing with sampling temperature. At temperature 0.8 its cumulative overlap fell to 1.2 s and gaps rose to 12.5 s, so a single “Moshi overlap” figure can be quoted either way. Most recent available data for dGSLM: 2022.
Summary: Conversational AI End-of-Turn Detection and Barge-In by the Numbers
| Metric | Value | Source |
|---|---|---|
| Listener start time in smooth handovers (before the turn ends) | Median 151 ms early | TurnBench, IEEE SLT 2026 |
| Turn transitions faster than 200 ms | 51% to 55% | Heldner and Edlund 2010, via Frontiers in Psychology 2015 |
| Response delay when the reply does not answer the question | +131.78 ms | Stivers et al., PNAS 2009 |
| Mean offset of responses coders judged on time, Japanese vs Danish | 36 ms vs 203 ms | Stivers et al., PNAS 2009 |
| Amazon Lex V2 default silence wait before ending the utterance | 600 ms | AWS Lex V2 documentation |
| Google median endpoint latency, separate vs joint model | 390 ms vs 270 ms | Bijwadia et al., Google, SLT 2022 |
| Robot turns that interrupted the user, silence-based baseline vs predictive model | 16.6% vs 6.9% | Skantze and Irfan, KTH, 2025 |
| Median robot response time, silence-based baseline vs predictive model | 2.7 s vs 1.5 s | Skantze and Irfan, KTH, 2025 |
| Precision predicting a true end of turn, acoustic detector vs end-to-end model | 67.3% vs 84.7% | Chang et al., Google, Interspeech 2022 |
| VAP accuracy, English-trained model on Japanese vs Japanese-trained | 59.43% vs 74.20% | Inoue et al., LREC-COLING 2024 |
| GPT-4o Realtime vs Nova Sonic, time to stop after user interruption | 0.23 s vs 2.25 s | Full-Duplex-Bench v1.5, 2025 |
| User interruptions ending with the user taking the floor, Moshi vs human speaker | 17.4% vs 63.6% | Arora et al., ICLR 2025 |
| Moshi theoretical / practical latency | 160 ms / 200 ms | Kyutai, 2024 |
| Overlap per minute, real Fisher calls vs cascaded system | 3.6 s vs 0.0 s | Nguyen et al., dGSLM, 2022 |
| Silent-gap share of conversation, Moshi vs cascaded system | 11.8% vs 32.4% | Arora et al., ICLR 2025 |
| Moshi takeover rate during user pauses, Candor set | 0.980 | Full-Duplex-Bench, 2025 |
Methodology and Sources
- Jiang et al., TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue, IEEE SLT 2026 and the live TurnBench leaderboard: human timing statistics and system recall, false positive rate and latency.
- Stivers et al., Universals and cultural variation in turn-taking in conversation, PNAS 2009: 10 languages, 101 conversations, polar questions.
- Levinson and Torreira, Timing in turn-taking and its implications for processing models of language, Frontiers in Psychology 2015: Switchboard analysis and review of Heldner and Edlund (Journal of Phonetics 2010).
- AWS, Configuring timeouts for capturing user input with a Lex V2 bot: default silence, start and barge-in settings.
- Bijwadia et al., Google, Unified End-to-End Speech Recognition and Endpointing for Fast and Efficient Speech Systems, SLT 2022.
- Chang et al., Google, Turn-Taking Prediction for Natural Conversational Speech, Interspeech 2022.
- Raju et al., Amazon, Two-Pass Endpoint Detection for Speech Recognition, 2024.
- Skantze and Irfan, KTH, Applying General Turn-taking Models to Conversational Human-Robot Interaction, 2025.
- Ekstedt and Skantze, TurnGPT, Findings of EMNLP 2020.
- Inoue et al., Multilingual Turn-taking Prediction Using Voice Activity Projection, LREC-COLING 2024 and Inoue et al., Real-time and Continuous Turn-taking Prediction Using Voice Activity Projection, IWSDS 2024.
- Lin et al., Full-Duplex-Bench, ASRU 2025 and Lin et al., Full-Duplex-Bench v1.5, 2025 (version 3, January 28, 2026).
- Arora et al., Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics, ICLR 2025.
- Défossez et al., Kyutai, Moshi: a speech-text foundation model for real-time dialogue, 2024.
- Nguyen et al., Generative Spoken Dialogue Language Modeling (dGSLM), 2022, published in TACL 2023.
- LiveKit, Using a transformer to improve end of turn detection, December 20, 2024: vendor claims, labeled as such.
- Data watch: the Moshi paper describes a 230 ms average response time across 10 languages, while the PNAS study reports a 208 ms mean, so cross-language averages differ by source. Overlap prevalence also varies by method: about 30% of floor transfers in Levinson and Torreira’s Switchboard analysis versus about 40% of transitions in Heldner and Edlund. Google’s conference abstract lists median and 90th-percentile reductions of 130 ms (33.3%) and 160 ms (22.2%), while the arXiv tables used here give 120 ms (30.8%) and 170 ms (23.0%). The TurnBench leaderboard is live and accepts external submissions, so rankings will move; figures above were read on October 31, 2026, and we left out the self-submitted external entries, nine of which rank above Voice Activity Projection on recall. Full-Duplex-Bench v1.5 figures come from arXiv version 3. We found no primary public data on contact-center barge-in frequency.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the IEEE SLT 2026 proceedings publish the final TurnBench paper.