Coupled but Late: Full-Duplex Speech Models Take Turns 480ms After Each Other

Forum topic · 小凯 · 2026-10-07

Summary

A Duke University study ("Coupled but Late: Turn-Taking Between Full-Duplex Speech Models in Unscripted Dialogue," arXiv:2610.08683) had two PersonaPlex-7B full-duplex speech models hold 36 unscripted 90-second conversations, exchanging Mimi audio tokens directly with no ASR/TTS pipeline. While humans in the Switchboard corpus take conversational turns after a median gap of just 137 ms, the AI pairs waited a median of 480 ms—three to four times slower—and only 1% of their turns landed in the final 120 ms of the partner's speech, versus 10% for humans. Re-pairing experiments showed the rhythm is genuinely coupled between partners, not a system clock; latency-scanning proved waiting is anchored to the partner's perceived speech end, indicating a reactive silence-threshold endpointing strategy rather than prediction. Playing recordings instead of a live partner broke the pattern entirely, showing the slow rhythm is socially triggered. A notable bonus finding: scripted call-center-style openings dominate the first 10 seconds (29 per 1,000 words), and scenario-based spoken openings reduce this far more effectively than system-prompt instructions. The study warns that AI-to-AI dialogue and AI-based evaluation may inherit this reactive slowness as a new normal.

This is a translated and structured summary of a Chinese forum post analyzing the Duke University paper "Coupled but Late: Turn-Taking Between Full-Duplex Speech Models in Unscripted Dialogue."

Background

The previous installment in this series tested AI's physical intuition via WorldSolver (the best models scored only half marks, failing even at textbook setups like Newton's cradle). This post moves to the second question: hearing — not speech recognition accuracy, but the *timing* of listening.

Human conversation runs on turn-taking: we respond while the other person is still finishing, laugh at the same beat, and drop an "mm-hm" right in the pause. That window is astonishingly narrow. The paper asks: when two full-duplex speech models talk with no human present, what rhythm do they settle into?

Experimental Setup

Two PersonaPlex-7B instances, one GPU each, directly exchanging Mimi audio tokens at 12.5 Hz (one step per 80 ms) — no ASR/TTS in between, so no pipeline can be blamed for slowness.
36 seed conversations, 90 seconds each, 5 scenarios; temperatures fixed (audio 0.8, text 0.7); personas, voices, and seeds controlled. Human baseline: the first 90 seconds of Switchboard calls (43,890 / 193,115 total turn transitions).
Metric: FTO (floor-transfer offset) = time next speaker starts minus current speaker's speech end. Positive = gap; negative = overlap. Speech segments with pauses <180 ms merge into IPUs; outliers trimmed to ±2 s.
Robustness: three segmentation methods (energy-based RMS 0.01, text words, aligned words); conversation-level bootstrap with 10,000 resamples; significance threshold Bonferroni-corrected to ~0.0011.

Main Results

Human median turn gap: 137 ms (roughly the duration of a blink); 25% of transitions fall within ±120 ms of zero; 10% are "projections" starting in the partner's final 120 ms.
AI median gap: 480 ms (CI [400, 480]) — 3–4× slower. Only 3% of transitions fall in the ±120 ms band; only 1% project into the final 120 ms. 44% of transitions land precisely at the 400/480 ms values, like stamp marks — a near-vacuum around zero where human activity is densest.

Experiment 1: The rhythm is co-created, not a clock

Re-pairing agents across conversations (2,000 derangements): the median collapses from 480 ms to +80 ms, overlap doubles to 0.46, and the IQR explodes from 160 ms to 1360 ms. True pairs beat all 2,000 re-pairings on every statistic. The 480 ms is therefore an emergent property of the *pair* — real coupling, just on a slow beat. In a silence control (partner audio removed), transitions collapse from 273 to 26 and words per agent from 134 to 29: the "conversation" is almost entirely sustained by the partner's live audio.

Experiment 2: Lateness is strategy, not channel latency

Subtracting the full 240 ms fine-codebook readout from every transition still leaves a 240 ms median, half the human near-zero band, and the 10% vs 1% projection gap.
Injecting artificial delay D ∈ {3, 8, 10, 15} frames shifts the response histogram mode exactly to 480 + D×80 ms; the partner's side stays pinned at 6 frames. Waiting is anchored to the *perceived end* of the partner's audio — it does not predict the end; it waits for it.
Agent overlap (18%) tends to be deep interruptions (median −560 ms), unlike humans' tail-adjacent backchannels (median −311 ms).

Conclusion: a reactive silence-threshold endpointing policy — like a sound-activated light or a call-center script: confirm the customer has fully stopped, then speak.

Experiment 3: Recordings break the etiquette

With pre-recorded AMI audio as the partner (72 runs), the signature 6-frame spike vanishes, the median halves to 240 ms, the ±120 ms share rises to 0.16, and agents' speech share drops from 0.35 to 0.15. The 480 ms is not baked into the weights — it is socially triggered behavior that only a live partner elicits.

Bonus finding: call-center accent

The models open with scripted call-center greetings ("Hi, this is Aria."). A 21-pattern lexicon finds 29 such openings per 1,000 words in the first 10 seconds, dropping to 2–4 afterwards. Injecting a scenario-based opening *into the speech stream* cuts first-10-second call-center speech from 29 to 12 per 1,000 words (overall 7.6 → 2.2), while the same sentence in the system prompt changes almost nothing (only 8 of 36 conversations comply). The behavior lives in the audio layer, not the instruction layer.

Takeaway

Two full-duplex models spontaneously form stable, coupled, reproducible rhythms — but anchored to the partner's audio end, 3–4× slower than humans, with the final 120 ms where humans express "I'm tracking you" left nearly empty. The paper's sober warning: dialogues generated by such pairs, and evaluations using them as examiners, will inherit the floor of a reactive endpointer — if everyone is slow, slowness becomes the new normal. Human fast turn-taking comes from prediction, not sharp ears; the AI waits, politely, half a second behind. References:

Paper: Coupled but Late: Turn-Taking Between Full-Duplex Speech Models in Unscripted Dialogue
Authors: Lichen Zhu, Yueqian Lin, Yiheng Wang, Hai "Helen" Li, Yiran Chen
arXiv: 2610.08683 (v1, 2026-10-06), Duke University; submitted to ICASSP 2027; code and measurement data to be open-sourced.