The 200-Millisecond Miracle: Why Taking Turns Is the Hardest Easy Thing Robots Do

Humans swap the conversational floor faster than they can plan a single word, and that small fact explains why so many robots still feel like walkie-talkies.

A system nobody was taught

Watch any two people chatting over coffee and you will see a feat of coordination that would impress an air-traffic controller. One person stops, the other starts, and the handover is so smooth that neither notices it happening. Nobody learns the rules at school, yet almost everyone follows them.

In 1974 the sociologists Harvey Sacks, Emanuel Schegloff and Gail Jefferson published what became the founding paper of conversation analysis. Working from recordings of ordinary talk, they proposed a compact model of how turns are allocated. Speech is built from units that a listener can recognise as potentially complete, and each such point is a place where the floor may change hands. At that moment the current speaker can pick the next one, someone else can claim the turn, or the speaker can simply carry on. The striking part is that this local, moment-by-moment machinery is enough to explain why talk usually involves one person speaking at a time, with little silence and little overlap.

Fast everywhere, but not equally fast

For decades it was easy to suspect that this tidy picture reflected English or a handful of European languages. In 2009 Tanya Stivers and an international team led from the Max Planck Institute for Psycholinguistics tested it properly. They analysed video recordings of informal conversation in ten languages spread across five continents, from Papua New Guinea to Namibia to Mexico, focusing on how quickly people answered yes-or-no questions.

Every language showed the same two tendencies: speakers avoided talking over each other and kept silences short. The languages did differ, but only modestly, with average response times falling within about a quarter of a second of the overall mean. Later summaries of the data put the extremes at roughly 7 milliseconds for Japanese and close to half a second for Danish. So the stereotype of the taciturn Scandinavian has a grain of truth, but even the slowest group answers faster than you can say the word 'pause'. The typical gap across many corpora sits around 200 milliseconds, the length of a single syllable.

The prediction puzzle

Here is where things get strange. Stephen Levinson and Francisco Torreira pointed out in 2015 that this speed is incompatible with how long it takes to produce speech. Laboratory studies show that naming a single picture takes roughly 600 milliseconds from seeing it to starting to speak, and building a short sentence takes longer still, sometimes more than a second. If listeners waited for the other person to finish before planning a reply, conversation would be full of long, awkward gaps.

The conclusion is that listeners must be doing two things at once. While the other person is still talking, they are working out what is being said, guessing how and when the turn will end, and already assembling their own response. In a 2016 review Levinson argued that this overlap of comprehension and production is one of the central demands that language makes on the brain, and that it may even have shaped how languages are structured. He also noted that turn-taking appears in infants before they have much language, and that versions of it exist across other primates, hinting that it is an old foundation rather than a late add-on.

Why robots stumble

Now consider how a typical spoken dialogue system works. It listens until the user goes quiet, waits for a silence threshold to pass (often somewhere between half a second and a second), sends the audio to a speech recogniser, passes the text to a language component, generates a reply and finally synthesises speech. Each stage waits for the one before. The result is a machine that is reactive where humans are predictive.

This creates two failure modes, and they pull in opposite directions. Set the silence threshold short and the robot barges in whenever the user pauses to think mid-sentence. Set it long and every exchange gains a dead patch that people read as confusion, rudeness or a crash. In his 2021 review of turn-taking in conversational systems, Gabriel Skantze of KTH described exactly this pattern: systems that interrupt too readily and respond too slowly, often both in the same conversation.

Robots add complications of their own. A physical body invites people to talk to it as they would to a person, which raises expectations about timing. Microphones on a moving head pick up motor noise. And people pause for many reasons: to search for a word, to breathe, to let a point land. A silence is not a signal in itself; its meaning depends on everything around it.

Learning to anticipate

The lesson from the human research is not that robots need faster processors. It is that they need to stop treating turn-taking as a matter of detecting silence and start treating it as a matter of prediction. People use grammar, intonation, gaze and the logic of the exchange (a question calls for an answer) to forecast where a turn is heading. A robot that wants to keep up must do the same.

That shift, from waiting to anticipating, is where the most interesting research in conversational robotics now sits. The next article in this series looks at the models trying to make it happen.

References

1. Sacks, H., Schegloff, E. A. and Jefferson, G. (1974). A simplest systematics for the organization of turn-taking for conversation. Language, 50(4), 696-735. Link
2. Stivers, T., Enfield, N. J., Brown, P., et al. (2009). Universals and cultural variation in turn-taking in conversation. Proceedings of the National Academy of Sciences, 106(26), 10587-10592. Link
3. Levinson, S. C. and Torreira, F. (2015). Timing in turn-taking and its implications for processing models of language. Frontiers in Psychology, 6, 731. Link
4. Levinson, S. C. (2016). Turn-taking in human communication: origins and implications for language processing. Trends in Cognitive Sciences, 20(1), 6-14. Link
5. Skantze, G. (2021). Turn-taking in conversational systems and human-robot interaction: a review. Computer Speech and Language, 67, 101178. Link

Feature drafted with AI from the published research listed above, for social-robotics.com.