📄 social_science

Quantifying sources of variation in turn-taking: How much do demographics and personality explain?

This study analyzes over 12,000 dyads to demonstrate that while demographics and personality explain a small portion of turn-taking variation, the majority of timing differences are driven by unique speaker, interlocutor, and dyadic interactions that cannot be fully predicted by individual attributes alone.

Original authors: Julio Cesar Cavalcanti, Plinio A. Barbosa, Sandra Madureira, Tony Berber Sardinha, Gabriel Skantze

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Julio Cesar Cavalcanti, Plinio A. Barbosa, Sandra Madureira, Tony Berber Sardinha, Gabriel Skantze

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Dance of Conversation

Imagine you are at a busy party. You're trying to have a conversation with a friend, but the room is loud. To keep the chat flowing, you and your friend have to perform a high-speed, invisible dance. You don't just wait for the other person to stop talking and then start; you have to predict exactly when they will finish so you can jump in at the perfect moment. If you jump in too early, you interrupt (an "overlap"). If you wait too long, there's an awkward silence (a "gap").

Scientists call this turn-taking. It's the rhythmic back-and-forth that makes human conversation feel smooth rather than like a game of ping-pong where the ball hits the floor every time. For a long time, researchers have debated why we dance this way. Is it because of who we are? Maybe you are naturally a fast talker because you are extroverted, or maybe you speak slowly because of your age or gender. Or, is it something that happens between two people? Maybe the dance isn't about your individual steps, but about how well you and your specific partner sync up, regardless of who you are individually.

This question matters because if we want to build robots that talk like humans, or apps that help people communicate better, we need to know: do we need to program the robot with your personality, or do we need to teach it how to dance with you specifically?

The Great Conversation Experiment

In this study, a team of researchers decided to settle this debate by looking at a massive amount of real-world data. They didn't just watch a few people chat; they analyzed over 12,000 conversations from two huge databases. One database, called Fisher, contained telephone calls between strangers. The other, CANDOR, had video calls between strangers. Because many of the people in these databases talked to multiple different partners, the researchers could play a clever game of "spot the pattern."

They asked: If Person A talks to Person B, and then Person A talks to Person C, does Person A talk the same way both times? (This would mean it's all about the Actor). Or, does Person A talk differently depending on whether they are talking to Person B or Person C? (This would mean it's about the Partner). And finally, is there a special "magic" that only happens when Person A and Person B are together, a pattern that disappears if either of them talks to someone else? (This is the Dyad, or the pair).

To measure this, they looked at four specific "dance moves":

Floor Transfer Offset (FTO): The exact milliseconds between when one person stops and the other starts.
Overlap: How much time people spend talking over each other.
Pauses: How much silence happens while a person is holding the floor.
Speech Activity: How much of the total conversation time a person actually gets to speak.

The Surprising Results: It's Complicated!

The researchers found that there is no single answer. The "dance" depends entirely on which move you are watching.

The "Dyad" Wins the Timing Race
When it came to the split-second timing of turn-taking (the FTO), the most important factor was the pair itself. The researchers found that the specific chemistry between two people explained the most variation in their timing. Even after accounting for who the speakers were and who their partners were, the unique "vibe" of that specific pair still held the biggest sway. It's as if every couple has their own secret rhythm that you can't predict just by knowing the individuals. In the telephone data, the pair explained 54% of the timing differences; in the video data, it was 34%. This suggests that the "clock" of a conversation is set by the two people dancing together, not just by the individual dancers.

The "Actor" Leads the Silence
However, when it came to Pauses (silence while holding the floor), the individual speaker was the boss. The researchers found that how much you pause is mostly about you. If you are a person who pauses a lot, you will pause a lot no matter who you are talking to. This was the most "actor-led" measure, meaning your personal style dominates the silence.

The "Speech Activity" is a Team Effort
For Speech Activity (who gets to talk the most), it was a mix. Your own personality and habits were the biggest factor (the Actor), but your partner mattered a lot too. If you talk to a dominant partner, you might talk less, even if you are usually a chatterbox. The researchers found that while your own traits explained the most, your partner's identity still accounted for a significant chunk of the variation.

The "Overlap" is a Shapeshifter
The amount of talking over each other (Overlap) was the most confusing. In the video conversations, it seemed to depend mostly on the speaker. But in the telephone calls, it depended mostly on the listener. This suggests that how we interrupt or overlap might change based on the medium (video vs. audio) and how we interpret the other person's cues.

Do Your Demographics Matter?

You might be wondering: "So, does being a man, a woman, young, old, or having a certain personality change how I talk?"

The answer is: Yes, but not as much as you might think.

The researchers tested if things like Sex, Age, Education, and Personality could predict these patterns. They found that these factors did matter, but they only explained a small slice of the puzzle—roughly 6% to 16% of the total variation.

Sex was the most consistent predictor. Men tended to have slightly longer gaps between turns and talked over each other slightly less than women.
Age and Personality (specifically extraversion) also played a role, especially in how much people talked.
However, even with all these details, the researchers couldn't predict a conversation just by looking at a person's profile. The "magic" of the specific pair and the unique context of the moment still ruled the day.

What This Means for the Future

The study concludes that we are more than just our demographic profiles. We are not just "extroverted men" or "older women" with fixed conversation styles. We are dynamic dancers who change our steps depending on who we are dancing with.

For the robots and AI systems of the future, this is a big clue. If you want a robot to talk naturally, you can't just program it with the user's personality. You have to teach it to listen to the pair. The robot needs to learn that this specific human talks differently with this specific robot than they do with their mom or their boss. The timing of a conversation isn't just a solo performance; it's a duet, and the music is often written by the two people playing it together, not by the sheet music in their pockets.

Technical Summary: Quantifying Sources of Variation in Turn-Taking

Problem Statement
Human conversation relies on precise temporal coordination, yet the distribution of timing variation remains unclear. A fundamental debate exists in the behavioral sciences regarding whether turn-taking timing is a consistent characteristic of the individual (driven by demographics and personality) or an emergent property of the dyadic system (driven by the specific pairing). Traditional approaches often attribute variation to "who the speaker is," while interactionist theories suggest timing arises from mutual entrainment. Historically, disentangling idiosyncratic traits from partner-associated elicitation and unique pair chemistry has been difficult due to small samples and narrow speaker pairings. This study seeks to quantify how much demographics and personality explain turn-taking variation compared to recurring speaker tendencies, interlocutor tendencies, and dyad-specific structures.

Methodology
The authors analyzed over 12,000 dyads from two distinct corpora:

CANDOR: A multimodal, video-mediated corpus of naturalistic online conversations (minimum 25 minutes).
Fisher: An audio-only telephone corpus of assigned-topic conversations (approx. 10 minutes).

Data Structure:
More than half of the participants in both samples appeared with multiple interlocutors. This recurrence allowed the use of mixed-effects models to separate variation into three random-intercept components:

Idiosyncrasy (Actor): Variation recurring with the focal speaker across different interlocutors.
Allosyncrasy (Partner): Variation in focal speakers' outcomes that recurs when a specific person serves as the interlocutor.
Dyadosyncrasy (Dyad): Additional variation associated with the specific unordered pair.

Metrics:
Four turn-taking measures were examined:

Floor Transfer Offset (FTO): Mean temporal offset between turns (gaps or overlaps).
Overlap: Ratio of a speaker's total overlap time to their total speaking time.
Pauses: Ratio of within-turn silence to total active time.
Speech Activity: Ratio of total speaking time to total interaction time.

Predictors:
Fixed effects included demographic and dispositional attributes (Sex, Age, Education, Race/Ethnicity, Personality, Social Status, Dialect, Topic). The study employed actor-oriented mixed models where actor and partner roles reverse in the second observation of a conversation. Model fit was assessed using marginal $R^2$ (fixed effects only) and conditional $R^2$ (fixed + random effects). Importance was determined via full-minus-restricted ablation (removing specific intercepts or predictor blocks) and variance decomposition.

Key Results

Sources of Variation:
- FTO: Showed the clearest dyad-associated structure. Removing the dyad intercept caused the largest loss in conditional $R^2$ in both corpora. In Fisher, the dyad accounted for 54% of total variance, compared to 11% (actor) and 8% (partner).
- Pauses: Were consistently actor-led. Actor variance represented 31–37% of total variance, significantly higher than partner or dyad shares.
- Speech Activity: Was actor-dominant but included appreciable partner-associated variation. Actor shares were 37% (CANDOR) and 48% (Fisher), with partner shares at 21% and 13% respectively. A shared dyad intercept could not be retained in the primary model due to the structural interdependence of floor shares.
- Overlap: The leading source of variation depended on the corpus and normalization. In CANDOR, it was actor-led; in Fisher, it was partner-led.
Predictive Power of Demographics and Personality:
- Fixed effects (demographics and personality) accounted for approximately 6–16% of total outcome variance.
- Sex and Age were the most consistent demographic predictors across metrics and corpora.
- Personality significantly improved the expanded models for all four outcomes in CANDOR, with the strongest effect on Speech Activity ( $\Delta R^2_m = .062$ ). In the basic models (containing only Sex, Age, and Education), Personality was not a significant predictor for FTO ( $p=.116$ ) or Pauses ( $p=.244$ ); it became significant only when the model was expanded to include additional covariates (e.g., Race/Ethnicity, Status), where it reached marginal significance for Pauses ( $p=.037$ ).
- Removing the dyad intercept resulted in the largest conditional $R^2$ loss for FTO, whereas removing the actor intercept caused the largest loss for Pauses and Speech Activity.
Corpus Differences:
- CANDOR speakers exhibited longer within-turn pauses than Fisher speakers. The authors note this may reflect multimodal coordination (visual cues) in video calls, though the study cannot isolate modality from other confounding factors (period, task, technology).
- Fisher's partner-led Overlap and larger dyad shares for FTO/Pauses suggest more stable pair-specific acoustic timing patterns in audio-only settings.

Significance and Claims
The paper claims to distinguish three specific types of variation:

Idiosyncrasies: Variation recurring with the focal speaker.
Allosyncrasies: Variation recurring with a person serving as an interlocutor.
Dyadosyncrasies: Variation associated with the particular dyad.

The study concludes that no single component dominates every turn-taking measure. While demographics and personality provide a baseline, they do not exhaust the variation associated with speakers, interlocutors, and dyads. Specifically:

FTO is primarily a property of the dyadic system.
Pauses are primarily a property of the individual.
Speech Activity is a mix of individual dominance and interlocutor influence.

The authors caution that while the dyad component captures shared shifts in timing, it does not definitively prove mutual accommodation versus unilateral adaptation. They suggest that conversational systems (e.g., human-robot interaction) should not rely solely on user-level averages for timing policies but must account for interaction-specific calibration. The findings support a multi-parametric view where the "universal clock" of turn-taking is calibrated differently depending on the specific metric being measured.