Across 10 languages on five continents, the average answer to a question starts just 208 milliseconds after the question ends (Stivers et al., PNAS 2009). That is far shorter than the roughly 600 ms it takes a person to begin naming a single picture (Levinson and Torreira, Frontiers in Psychology 2015), so speakers must plan their reply while the other person is still talking. Video calls break this rhythm: pairs who switched turns in 135 ms face to face needed 487 ms over Zoom (Boland et al., Journal of Experimental Psychology: General 2022). This roundup covers human conversation; for how software decides that a speaker has finished, see our end-of-turn detection and barge-in statistics. We aggregated data from PNAS, the Max Planck Institute for Psycholinguistics, the Journal of Pragmatics, Interspeech, Child Development, the ITU, and more than a dozen other peer-reviewed sources.
TL;DR
- The mean response to a question starts +208 ms after it ends, from +7 ms in Japanese to +469 ms in Danish (Stivers et al., PNAS 2009).
- About 40% of all between-speaker intervals in three conversation corpora are overlaps (Heldner and Edlund, Journal of Phonetics 2010).
- 10,950 US English phone conversations show an average gap of 0.300 s between turns (Cavalcanti and Skantze, Interspeech 2025).
- Beginning to name a single picture takes about 600 ms (Levinson and Torreira, Frontiers in Psychology 2015, citing Indefrey and Levelt 2004).
- Listeners start planning within half a second after the answer can be retrieved, sometimes about 2.4 s before the question ends (Bögels, Magyari and Levinson, Scientific Reports 2015).
- Dispreferred responses clearly outnumber preferred ones only after gaps of 700 ms or more (Kendrick and Torreira, Discourse Processes 2015).
- Faster average response times predict stronger felt connection, b = −0.28, P < 0.001 (Templeton et al., PNAS 2022).
- Infants overlap just over a third of their vocalizations with their mothers at 3 to 5 months, roughly 20% at 18 months (Hilbrink, Gattis and Levinson, Frontiers in Psychology 2015).
- Across 26 studies, infants take on average circa 1 s to respond (Nguyen et al., Child Development 2022).
- Turn transitions rose from 135 ms face to face to 487 ms over Zoom for the same pairs (Boland et al., Journal of Experimental Psychology: General 2022).
- One-way delay below 150 ms gives essentially transparent interactivity; 400 ms is the planning ceiling (ITU-T Recommendation G.114, 2003).
1. Average Gap Between Turns Across Languages
The headline finding of the largest cross-language comparison is not that cultures differ, but how little they differ. In Stivers and colleagues’ study of 10 languages, the most common response offset was exactly 0 ms, and every language mean fell within about 250 ms either side of the pooled mean. Danish, the slowest language in the sample, still answered in under half a second on average (+469 ms). The authors suggest that ethnographic descriptions of strikingly “slow” or “fast” conversational cultures reflect a sensitivity to tempo differences of only a few hundred milliseconds.
The measure used here is the floor transfer offset (FTO): the time from the end of one speaker’s turn to the start of the next speaker’s turn, where a negative value means the next speaker started early, in overlap. The per-language means, as tabulated by Levinson and Torreira (Table 1), were Tzeltal 67 ms, Yélî Dnye 71 ms, Dutch 108 ms, Korean 182 ms, English 236 ms, Italian 309 ms, Lao 419 ms and ‡Ākhoe Hai‖om 423 ms. Large English telephone corpora land in the same band: a mean of 187 ms on Switchboard and 0.300 s on Fisher.
| Metric | Value | Source |
|---|---|---|
| Size of the 10-language comparison | 350 consecutive questions per language, 101 conversations in total | Stivers et al., PNAS 2009 |
| Mean time from end of question to start of response, all 10 languages pooled | +208 ms | Stivers et al., PNAS 2009 |
| Most frequent (mode) and median response offset, all 10 languages pooled | Mode 0 ms, median +100 ms | Stivers et al., PNAS 2009 |
| Fastest vs slowest language, by mean response offset | Japanese +7 ms vs Danish +469 ms | Stivers et al., PNAS 2009 |
| Mean and median floor transfer offset, Switchboard English telephone corpus (19,754 turn transitions, 348 conversations) | Mean 187 ms, median 168 ms | Roberts, Torreira and Levinson, Frontiers in Psychology 2015 |
| Average transition floor offset, Fisher US English telephone corpus (10,950 conversations, 11,684 speakers) | 0.300 s (SD = 0.228) | Cavalcanti and Skantze, Interspeech 2025 |
| Mean gap across the dyadic conversations used to define long gaps | 248 ms (s.d. = 598 ms) | Templeton et al., Philosophical Transactions B 2023 |
Context note: the 10-language figures cover only responses to questions, not every turn change, and most recent available data for that comparison is from 2009. The 2025 Fisher study adds a twist: demographics barely matter (sex explained about 3.2% of the variation and age 1.3%), while the specific pair of speakers mattered most, which the authors call “dyadosyncrasy” (Cavalcanti and Skantze, Interspeech 2025).
2. Overlap Rates and the Shape of the Gap Distribution
An average of around 200 ms hides a wide spread. Measured at every change of speaker, overlap is not a rare breakdown but a routine part of talk: in the corpora analysed by Heldner and Edlund, about 40% of all between-speaker intervals were overlaps. A between-speaker interval is the silence or the overlap at the moment one speaker stops and the other starts.
Yet overlap fills very little of the actual recording. In the Switchboard telephone corpus only 3.8% of the audio has both people talking at once, because most overlaps are brief. Heldner and Edlund also found that only 14 to 19% of gaps were shorter than 200 ms, so “no gap, no overlap” transitions are far less common than textbook descriptions suggest.
| Metric | Value | Source |
|---|---|---|
| Share of between-speaker intervals that are overlaps (Spoken Dutch Corpus / HCRC Map Task Corpus / Swedish Map Task Corpus) | 40.0% / 41.7% / 40.0% | Heldner and Edlund, Journal of Phonetics 2010 |
| Share of between-speaker intervals (gaps and overlaps) shorter than 500 ms | 70-82% | Heldner and Edlund, Journal of Phonetics 2010 |
| Share of between-speaker intervals that are either overlaps or gaps too short to notice | 55-59% | Heldner and Edlund, Journal of Phonetics 2010 |
| Share of Switchboard telephone audio with both speakers talking at once | 3.8% (77% one speaker only, 19.2% silence) | Levinson and Torreira, Frontiers in Psychology 2015 |
| Share of Switchboard floor transfers that start in overlap | 30.1% | Levinson and Torreira, Frontiers in Psychology 2015 |
| Backchannels overlapped by the next turn vs other turns overlapped (Switchboard) | 35% vs 28% | Roberts, Torreira and Levinson, Frontiers in Psychology 2015 |
| Overlapping speech as a share of all speech, 6 languages, about one hour each | 10.5% (Spanish) to 20.8% (Korean); English 12.6% | Liesenfeld, Lopez and Dingemanse 2023 |
| Mean turn duration in the same 6-language sample | 2,538 ms (Mandarin) to 4,620 ms (Spanish) | Liesenfeld, Lopez and Dingemanse 2023 |
Definitions differ, which explains why overlap shares do not match across rows: Heldner and Edlund count overlaps among all between-speaker intervals, Switchboard’s 30.1% counts floor transfers, and the 6-language sample in Liesenfeld, Lopez and Dingemanse measures overlap as a share of speech time. A backchannel is a short listener response such as “mm-hm” or “yeah” that does not claim the floor. Most recent available data for the Dutch, English and Swedish corpora is from 2010.
3. How Long It Takes to Plan a Reply
This is the central puzzle of turn-taking research. Producing even one word takes about 600 ms from seeing a picture to starting to speak, and a simple sentence takes about 1,500 ms, yet the 208 ms mean gap from section 1 is about a third of the single-word figure (208 / 600 = 0.35). The only way the numbers add up is if listeners begin planning their answer while still listening.
The EEG evidence supports that. In Bögels, Magyari and Levinson’s experiment, brain signatures of response planning appeared within half a second after the answer to a question became retrievable, in some questions about 2.4 seconds before the question ended. Answers that could be planned early came out faster: 950 ms minus 640 ms gives a difference of about 310 ms between the two conditions (our arithmetic from the reported means). The same budget problem now drives the design of voice software, as our time-to-first-audio benchmarks for voice agents show.
| Metric | Value | Source |
|---|---|---|
| Time from seeing a picture to beginning to name it (estimate) | about 600 ms | Levinson and Torreira, Frontiers in Psychology 2015, citing Indefrey and Levelt 2004 |
| Time before speech begins when describing a simple scene (e.g., boy kicking ball) in a sentence | about 1,500 ms | Levinson and Torreira, Frontiers in Psychology 2015 |
| Timing of an early EEG signature of preparing a response, relative to the end of the incoming turn | c. 1,200 ms before the end, about halfway through the turn | Levinson and Torreira, Frontiers in Psychology 2015, citing Magyari et al. 2014 |
| Participants and stimuli in the EEG question-answer experiment | 24 participants (main experiment), 94 Dutch question pairs | Bögels, Magyari and Levinson, Scientific Reports 2015 |
| Mean response time when the answer was available early in the question vs only at the end | M = 640 ms (mode around 400 ms) vs M = 950 ms (mode around 750 ms) | Bögels, Magyari and Levinson, Scientific Reports 2015 |
| How early the answer could be retrieved in the early-answer questions | around 2.4 seconds before the end of the question, on average | Bögels, Magyari and Levinson, Scientific Reports 2015 |
| Listeners in a Swedish experiment on turn-end cues, and the most disruptive change to the last 500 ms of a turn | N = 32; increasing the speech rate (p<.005) | Feindt et al., Interspeech 2023 |
Context note: in the Swedish Interspeech 2023 study, speeding up the end of a turn lengthened reaction times when the speaker was keeping the turn but shortened them when the speaker was yielding it, a sign that listeners read speech rate as a turn-end cue. Most recent available data for the naming latencies is from 2004 and earlier.
4. What a Gap Signals: Preference, Agreement and Connection
Because the default gap is so short, a slightly longer one carries meaning. Confirmation comes faster than its opposite: confirming answers were between 100 and 500 ms faster on average than non-confirming ones in all 10 languages of the comparison in section 1. In telephone corpora, Kendrick and Torreira found that refusals and other dispreferred responses clearly outnumbered preferred ones only once the gap reached 700 ms or more.
Speed also reads as rapport. In Templeton and colleagues’ 2022 PNAS study, shorter response times predicted stronger felt connection, and outside listeners rated edited fast versions of conversations as more enjoyable than slow ones. But the rule flips between friends: in a 2023 follow-up, friends had more long silences than strangers, and those silences were not awkward for them.
| Metric | Value | Source |
|---|---|---|
| Speed of confirming answers vs non-confirming answers to questions | 100 to 500 ms faster on average, in all 10 languages | Stivers et al., PNAS 2009 |
| Gap length at which dispreferred responses clearly outnumber preferred ones (195 responses in telephone corpora) | 700 ms or more | Kendrick and Torreira, Discourse Processes 2015 |
| Mean gap after a turn that contains a responding action vs after other turns (Switchboard) | 150 ms vs 202 ms | Roberts, Torreira and Levinson, Frontiers in Psychology 2015 |
| Effect of a person’s average response time on felt connection (66 participants, each in 10 ten-minute conversations) | b = −0.28, SE = 0.05, P < 0.001 | Templeton et al., PNAS 2022 |
| Mean gap in fast, original and slow edited versions of the same conversations, rated by 450 online listeners | 55.68 ms / 278.55 ms / 557.14 ms; response time inversely predicted perceived enjoyment, F(2, 2,351.9) = 49.44, P < 0.001 | Templeton et al., PNAS 2022 |
| Threshold used to define a long gap in conversation | greater than 2 s, roughly 3 standard deviations from the mean | Templeton et al., Philosophical Transactions B 2023 |
| Conversations analysed, and which group had more long gaps | 261 stranger dyads and 65 friend conversations; friends had more long gaps than strangers | Templeton et al., Philosophical Transactions B 2023 |
Context note: the Templeton findings come from one research group’s samples (66 participants in the 2022 study, 261 stranger pairs in the 2023 study), so treat them as evidence about those participants rather than about all speakers. A “dispreferred” response is one the question did not invite, such as declining an invitation or disagreeing.
5. How Children Develop Turn Timing
Turn-taking starts before words do, and the path is not a straight line. Babies at 3 to 5 months are quick but messy: in Hilbrink, Gattis and Levinson’s longitudinal study of 12 mother-infant pairs (8,555 turn transitions), infants produced just over a third of their vocalizations in overlap with their mothers. Around 9 and 12 months their gaps grew much longer, while mothers’ timing stayed stable.
An earlier report with overlapping authors described the same slowdown: children started on par with their mothers but showed significantly longer gaps at the 12-month sample (Casillas et al., SemDial 2013). Across 26 studies, the pooled evidence puts infant response latency at about one second, several times longer than the adult averages in section 1.
| Metric | Value | Source |
|---|---|---|
| Average turn transition in mother-infant protoconversation at about 3 months | about 1.5 s | Levinson and Torreira, Frontiers in Psychology 2015, citing Bateson 1975 |
| Infant vocalizations produced in overlap with the mother at 3, 4 and 5 months | just over a third | Hilbrink, Gattis and Levinson, Frontiers in Psychology 2015 |
| Mother turns produced in overlap with the infant at 3, 4 and 5 months | between 14 and 21% | Hilbrink, Gattis and Levinson, Frontiers in Psychology 2015 |
| Infant turns produced in overlap at 18 months | roughly 20%, similar to the mothers | Hilbrink, Gattis and Levinson, Frontiers in Psychology 2015 |
| Children overlapping the end of the mother’s turn before 5 months vs at 1 year and 8 months | nearly 40% of the time vs approximately 4% | Casillas et al., SemDial 2013 |
| Infant response latency pooled across 26 studies (78 estimates, 429 infants) | circa 1 s on average | Nguyen et al., Child Development 2022 |
| Change in infant response latency when adult latency rises by 1 s (e.g., 400 to 1,400 ms) | an increase of over 1 s | Nguyen et al., Child Development 2022 |
The last row is the most practical one: the 2022 meta-analysis found that infants mirror their partner’s tempo, so slower adults are answered by slower babies. Most recent available data for the longitudinal infant samples is from 2013 to 2015.
6. Video-Call Latency and Turn-Taking
Video calls add only tens of milliseconds of audio delay, but the effect on turn-taking is out of all proportion. In Boland and colleagues’ experiments, the same pairs who switched turns in 135 ms face to face took 487 ms over Zoom, a 352 ms increase (487 minus 135) against an estimated transmission delay of only 30-70 ms. In a yes/no task the gap between local and remote answers was 679 ms (976 minus 297). The authors argue the delay disrupts the timing mechanisms people use to launch a turn, not just the signal.
Field data from clinical video calls tells the same story. Seuren and colleagues found that around 100 ms of latency barely affected turn-taking, while at 700 ms participants struggled routinely, with unintended silences and talk-overs. Network engineers have long planned around similar limits, as covered in our VoIP jitter buffer and speech quality statistics.
| Metric | Value | Source |
|---|---|---|
| Yes/no answers to prerecorded questions: average latency for local vs remote (Zoom) responses | 297 ms vs 976 ms | Boland et al., Journal of Experimental Psychology: General 2022 |
| Turn transition time for the same pairs, face to face vs over Zoom | 135 ms vs 487 ms | Boland et al., Journal of Experimental Psychology: General 2022 |
| Estimated audio transmission delay in the Zoom experiments | 30-70 ms | Boland et al., Journal of Experimental Psychology: General 2022 |
| Video consultations analysed, and turn-taking problems found | 25 consultations (7 h 57 min); 130 noticeable silences and 172 overlaps | Seuren et al., Journal of Pragmatics 2021 |
| Latency at which turn-taking is barely affected vs routinely disrupted | around 100 ms vs 700 ms | Seuren et al., Journal of Pragmatics 2021 |
| One-way delay below which most applications have essentially transparent interactivity | below 150 ms | ITU-T Recommendation G.114, 2003 |
| One-way delay that should not be exceeded for general network planning | 400 ms | ITU-T Recommendation G.114, 2003 |
| One-way delay with no significant impact on perceived overall speech quality in highly interactive conversation (tested delays up to 1,600 ms) | up to 400 ms | Egger, Schatz and Scherer, Interspeech 2010 |
Outlier note: ITU-T G.114 itself warns that highly interactive tasks, including some speech and video conferencing, may be affected by delays below 100 ms, and the Interspeech 2010 test found that delay sensitivity rises with how interactive the conversation is. Quality ratings and turn timing are different measures, which is why a 400 ms delay can be rated acceptable and still produce talk-overs. Layered pipelines such as live interpretation add their own delay on top, as shown in our speech-to-speech translation latency statistics. Most recent available data for G.114 is the 2003 edition, which remains in force.
Summary: Conversational Turn-Taking Gap Duration by the Numbers
| Metric | Value | Source |
|---|---|---|
| Mean response offset after questions, 10 languages pooled | +208 ms | Stivers et al., PNAS 2009 |
| Share of between-speaker intervals that are overlaps | about 40% | Heldner and Edlund, Journal of Phonetics 2010 |
| Share of between-speaker intervals shorter than 500 ms | 70-82% | Heldner and Edlund, Journal of Phonetics 2010 |
| Mean and median floor transfer offset, Switchboard | 187 ms and 168 ms | Roberts, Torreira and Levinson, Frontiers in Psychology 2015 |
| Average gap, Fisher corpus (10,950 US English phone conversations) | 0.300 s | Cavalcanti and Skantze, Interspeech 2025 |
| Time from seeing a picture to beginning to name it | about 600 ms | Levinson and Torreira, Frontiers in Psychology 2015 |
| Overlapping speech as a share of all speech, 6 languages | 10.5% to 20.8% | Liesenfeld, Lopez and Dingemanse 2023 |
| Response time, answer available early vs late in the question | 640 ms vs 950 ms | Bögels, Magyari and Levinson, Scientific Reports 2015 |
| How early the answer could be retrieved in early-answer questions | around 2.4 s before the question ends | Bögels, Magyari and Levinson, Scientific Reports 2015 |
| Gap length at which dispreferred responses clearly dominate | 700 ms or more | Kendrick and Torreira, Discourse Processes 2015 |
| Effect of average response time on felt connection | b = −0.28, P < 0.001 | Templeton et al., PNAS 2022 |
| Group with more long gaps (over 2 s) | friends, not strangers | Templeton et al., Philosophical Transactions B 2023 |
| Infant vocalizations in overlap, 3 to 5 months vs 18 months | just over a third vs roughly 20% | Hilbrink, Gattis and Levinson, Frontiers in Psychology 2015 |
| Children overlapping mother’s turn end, before 5 months vs at 1 year 8 months | nearly 40% vs approximately 4% | Casillas et al., SemDial 2013 |
| Infant response latency across 26 studies | circa 1 s | Nguyen et al., Child Development 2022 |
| Turn transition time, face to face vs Zoom | 135 ms vs 487 ms | Boland et al., Journal of Experimental Psychology: General 2022 |
| Yes/no answer latency, local vs remote over Zoom | 297 ms vs 976 ms | Boland et al., Journal of Experimental Psychology: General 2022 |
| Video-call latency: barely affected vs routinely disrupted | around 100 ms vs 700 ms | Seuren et al., Journal of Pragmatics 2021 |
| One-way delay: transparent interactivity vs planning ceiling | below 150 ms vs 400 ms | ITU-T Recommendation G.114, 2003 |
Methodology and Sources
- Stivers et al., Universals and cultural variation in turn-taking in conversation, PNAS 2009
- Levinson and Torreira, Timing in turn-taking and its implications for processing models of language, Frontiers in Psychology 2015
- Roberts, Torreira and Levinson, The effects of processing and sequence organization on the timing of turn taking: a corpus study, Frontiers in Psychology 2015
- Heldner and Edlund, Pauses, gaps and overlaps in conversations, Journal of Phonetics 2010
- Cavalcanti and Skantze, Dyadosyncrasy, Idiosyncrasy and Demographic Factors in Turn-Taking, Interspeech 2025
- Liesenfeld, Lopez and Dingemanse, The timing bottleneck, arXiv 2307.15493, 2023
- Bögels, Magyari and Levinson, Neural signatures of response planning occur midway through an incoming question in conversation, Scientific Reports 2015
- Feindt et al., Cues to next-speaker projection in conversational Swedish, Interspeech 2023
- Kendrick and Torreira, The timing and construction of preference: a quantitative study, Discourse Processes 2015
- Templeton et al., Fast response times signal social connection in conversation, PNAS 2022
- Templeton et al., Long gaps between turns are awkward for strangers but not for friends, Philosophical Transactions of the Royal Society B 2023
- Hilbrink, Gattis and Levinson, Early developmental changes in the timing of turn-taking, Frontiers in Psychology 2015
- Casillas et al., Turn-timing in naturalistic mother-child interactions: a longitudinal perspective, SemDial 2013
- Nguyen, Versyp, Cox and Fusaroli, A systematic review and Bayesian meta-analysis of the development of turn taking in adult-child vocal interactions, Child Development 2022
- Boland, Fonseca, Mermelstein and Williamson, Zoom disrupts the rhythm of conversation, Journal of Experimental Psychology: General 2022
- Seuren, Wherton, Greenhalgh and Shaw, Whose turn is it anyway? Latency and the organization of turn-taking in video-mediated interaction, Journal of Pragmatics 2021
- ITU-T Recommendation G.114, One-way transmission time, 2003
- Egger, Schatz and Scherer, It takes two to tango: assessing the impact of delay on conversational interactivity on perceived speech quality, Interspeech 2010
- Data watch: the original 10-language paper reports Danish at +469 ms and Italian at +310 ms in its text, while the 2015 review table that reproduces it lists 468 ms and 309 ms; we use the original paper’s figures where we cite it directly. The 10-language comparison measures only answers to questions, so its means are not directly comparable with whole-corpus averages from Switchboard (187 ms) or Fisher (0.300 s). Overlap shares differ by definition (40% of between-speaker intervals, 30.1% of Switchboard floor transfers, 10.5-20.8% of speech time) and should not be compared as if they measured the same thing. The 6-language overlap sample covers only about one hour of talk per language. Most cross-language, child-development and planning data is older than three years; the newest additions are the Interspeech 2025 Fisher analysis and the 2022-2023 studies on Zoom, response speed and long gaps. The Zoom findings come from controlled experiments with an estimated, not measured per-call, transmission delay.
Last updated: November 8, 2026. We update this roundup quarterly, and the next refresh is expected when the Interspeech 2026 proceedings are indexed in the ISCA Archive.