Social AI: Turn-Taking Predictor

Group Dynamics

Speakers around the table 6

Anticipatory-gaze weight w_g 1.00

Silence-fairness weight w_s 1.00

Softmax temperature T 0.60

Live Readout

Speaker A0 has the floor

Turn #

Turns / min

0.0

Dominance (Gini)

0.00

Last selection entropy

0.00 bits

Candidate scores (next-speaker)

How it works

Each round, one avatar "holds the floor" for a randomly drawn turn length. While they speak, the other avatars are modelled as listeners whose simulated gaze drifts toward whichever candidate looks most likely to speak next — this is anticipatory gaze, a real and well-documented social signal: listeners statistically look more at the person about to take the floor in the seconds before a turn actually changes.

G_i = EMA of gaze received by listener i
S_i = silence_i / Σ silence_j   (fairness share)
score_i = w_g·G_i + w_s·S_i + noise_i
P(i next) = softmax(score_i / T)

w_g (gaze weight) — how much accumulated listener attention biases the next-speaker draw. High values let one visually "watched" avatar dominate the floor.
w_s (fairness weight) — how much time-since-last-spoke counts. High values push the model toward round-robin turn-taking.
Temperature T — low T makes the softmax nearly greedy (the top-scoring candidate almost always wins the floor); high T flattens it toward a uniform, unpredictable draw.
Dominance (Gini) — the Gini coefficient of cumulative speaking time across all avatars; 0 = perfectly equal airtime, 1 = one avatar has said everything.
Selection entropy — Shannon entropy (bits) of the softmax distribution at the moment the last turn changed; near 0 means the outcome was almost certain, higher means several avatars were realistically in contention.

Real-world relevance: this is the same class of model used by meeting-assistant AIs and smart-speaker arrays (e.g. active-speaker detection microphone arrays, Zoom's floor-tracking) to decide whose audio to foreground next, and by conversational-agent research modelling when a bot should interject in a multi-party chat.

Current speaker Gaze-accumulation beam

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.