Conversation turn-taking
A Socratic walk-through of conversation turn-taking — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #How do people coordinate who speaks next without usually agreeing on rules aloud?
Nobody taught you the rules of taking turns, and no two people agree them before talking. Yet a conversation between strangers runs with almost no collisions and almost no dead air — two independent systems, no shared clock, no referee, and still the handovers land. What information could each speaker be using that the other reliably provides?
Reasoning it through
REASONING #Begin with the obvious candidate: perhaps we wait for silence, then start. Measure the silences and that story fails. In recordings of ordinary conversation across many unrelated languages, the gap between one speaker finishing and the next beginning clusters around a couple of tenths of a second — roughly 200 milliseconds, with overlaps and longer gaps on either side. Now compare how long it takes to prepare speech: laboratory estimates from picture-naming put the journey from intention to articulation at something like 600 milliseconds for a single word, and a sentence needs more. The arithmetic refuses the waiting story. Begin planning when the other person stops and you arrive late every time.
So the reply must be built while the other person is still talking, and launched at a moment predicted in advance. Predicted from what? Falling pitch and fading loudness help, but consider how you know a question is nearly over: usually from its content and grammar. "So what time does the…" already tells you the shape of the words remaining and where the structure completes. Experiments that strip a turn's words while keeping its intonation, and the reverse, find listeners predict endings far better from the lexical and syntactic content than from melody alone; prosody, gaze and a settling gesture seem to refine the timing rather than carry it.
Then a second problem, separate from when: who? If everyone able to reply launched at the projected ending, every handover would pile up. Watch what settles it. Sometimes the speaker chooses — a name, a direct question, a look held on one person — and that person's turn is effectively reserved, so their silence becomes conspicuous rather than neutral. When nobody has been chosen the floor is open, and in practice the first to start takes it. If nobody starts, the current speaker may simply continue, which is why an unanswered question so often grows an extra clause on its end. Conversation analysts describe these as an ordered set of options available at each point of possible completion, and the ordering is what makes being selected and staying silent feel loud while not being selected and staying silent feels like nothing.
Does that cover interruption? Partly. Much overlap is not competitive at all — a listener starting slightly early on a correctly predicted ending, or saying "mm" where it claims no floor — and real collisions tend to resolve within a syllable or two. The system is not a rule against simultaneous talk; it is a set of expectations precise enough that collisions are rare and get noticed when they happen.
The analogy
THE ANALOGY #Think of a relay handover. The incoming runner does not stand still waiting for the baton; she is already at speed, timing her launch to the approach of the runner behind, so the exchange happens at pace rather than from a standstill. Everything hinges on starting before the thing you are waiting for arrives, using its approach as the cue.
the relay's order is fixed in advance and the handover zone is painted on the track, so the runner never has to work out whether it is her turn or where the zone begins. In conversation both must be inferred from the content of the turn itself — which is exactly where the interesting machinery sits, and a problem the runner does not have.
Clarifying the model
THE MODEL #Two misconceptions are worth dislodging. First, turn-taking is not politeness. Polite cultures and blunt ones need a solution to the same coordination problem; what varies is mainly the tolerated size of gaps and overlaps. Silence that reads as thoughtful in one community reads as rejection in another — just what you would expect if the timing is a shared expectation rather than a law.
Second, the rules are not stored anywhere as rules. Speakers have the ability to project completion and a strong sense of what a missing reply means, and those two generate the orderly pattern observers later write down. What stays unsettled is the mix: how much of the prediction rides on grammar, how much on prosody, how much on the sheer predictability of the content.
A picture of it
THE PICTURE #How to readRead downward as time. The two notes on the listener's side are the heart of it — planning begins during the other person's turn, not after it, which is what makes a two-tenths-of-a-second handover possible at all. The three branches are what can happen at the projected ending, in the order the system tries them: a selected speaker replies, or else the floor is open to whoever starts, or else the original speaker simply keeps going.
What became clearer
WHAT CLEARED #Turn-taking works not because we wait our turn but because we forecast it. Listeners use grammar and content to predict where a turn will finish, build their reply while it still runs, and settle who goes next by a small ordered set of expectations — selection, then first-starter, then continuation — none of which anyone states aloud.
Where to go next
ONWARD #- Why video calls feel harder: what a few hundred milliseconds of transmission delay does to projection.
- How repair works when the handover fails — the machinery of "sorry, go on".
- Whether the same timing profile holds in signed languages, where the channel is visual.
Key terms
TERMS #| Term | What it means |
|---|---|
| Transition-relevance place | a point where a turn could be complete, and so where a handover legitimately may occur. |
| Turn projection | a listener's running prediction of where the current turn will finish, used to time their own start. |
| Backchannel | a short response such as "mm" or "right", produced without any claim on the floor. |
Every term the collection defines is gathered in the glossary.