Full-Duplex Speech Models Take the Floor When Asked, Not When Needed
Linkai Peng, Baorian Nuchged, Kaiqi Fu, Yuyang Yao
cs.CL · cs.HC
Quick verdict
Argues full-duplex speech models speak when addressed, not when content demands; well-controlled design, but lacks significance tests and judge validation.
Summary
This paper asks whether full-duplex speech models take the floor on their own when the content of a user's utterance demands it, as human listeners do when correcting false claims or warning of danger. The authors built 40 context-matched English monologues in which only a trigger utterance varies within a topic, defined ten conditions based on turn-allocation rules, and compressed inter-word pauses to under 0.12 s to separate the reason to speak from the opportunity created by silence. Across seven configurations from five model families (Moshi, PersonaPlex, VoiceChat-11B, MiniCPM-o 4.5, Raon-SpeechChat; 2,000 trials each), direct questions and inserted silence reliably raise speech onset, whereas false facts, verbatim repetition, and hazards stay within 0.06 of the neutral baseline. Text-token probability probes in Moshi and PersonaPlex confirm no raised tendency to speak after false facts. Granting extra pauses or explicit permission does not close the gap, and in a turn-release experiment only 14-15% of false-fact replies and 4-7% of hazard replies actually intervene. Main limitations: purely descriptive statistics without significance tests, an unvalidated LLM judge for reply quality, and synthetic TTS stimuli in English only.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 2
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Across five full-duplex model families, speech onset responds to being addressed and to silence, while false facts, verbatim repetition, and hazards stay within 0.06 of the neutral baseline (Table 2).
- Frame-level text-token probabilities in Moshi and PersonaPlex show no elevated tendency to speak after false facts (2 s means -0.03 and -0.02 vs Neutral), while questions and silence raise it (Sec. 4.2).
- Adding opportunity (inserted silence, natural pauses) or explicit permission instructions does not produce a consistent content-specific increase in onset (Table 3).
- Even when handed the floor, only .14-.15 of non-empty false-fact replies challenge the claim and .04-.07 of hazard replies warn of danger (Table 4).
Strengths
Novel question grounded in conversation analysis
Recasts model-initiated speech as Sacks/Schegloff/Jefferson self-selection (Sec. 1-2) and is positioned, credibly, as the first protocol comparing multiple triggers in matched contexts while controlling pauses, against FD-Bench, FLEXI and Instruct-FD.
Rigorous opportunity control with sharp negative controls
Compressing inter-word gaps to <=0.12 s (Sec. 3.1), plus rhetorical-question and turn-end conditions (Sec. 4.1), cleanly show models react to surface cues (being addressed, actual silence) rather than question form or yielding words.
Triangulated measurement across three levels
Breadth and transparency
Weaknesses
No inferential statistics anywhere
Unvalidated LLM judge carries the headline numbers
Ecological validity of compressed-pause TTS stimuli
Cross-family measures are not commensurable
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This Interspeech-style paper (five model families, seven configurations, 2,000 trials per configuration) asks whether full-duplex speech models exercise self-selection - taking the floor because the content demands it (false claim, missing word, imminent hazard) - as opposed to reacting to surface turn-allocation cues (being addressed, silence). The central claim, supported by Table 2, is that onset follows conversation structure, not content: direct questions raise onset by +0.08 to +0.24 over Neutral and 1.5 s of inserted silence lifts it to .83 (PPlex) or .98 (MiniCPM-o), while false facts, verbatim repetition, and hazards never deviate from Neutral by more than 0.06. Two elegant controls make this interpretable: rhetorical questions (question form without addressing) and turn-end phrases ("that's all from me" without stopping) have almost no effect, pinning the behavior to addressing and actual silence. A secondary claim (Table 4) is that even when handed the floor, intervention is rare: .14-.15 of false-fact replies challenge the claim, and only .04-.07 of hazard replies warn.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does it mean that full-duplex speech models take the floor when asked, not when needed?
The paper shows these models reliably start speaking when directly addressed or when the user falls silent, but they almost never speak on their own initiative to correct a false statement or warn about a danger - behaviors human listeners perform spontaneously without being asked.
Which speech models were evaluated and how many trials were run?
Seven configurations across five families: Moshi (Moshiko/Moshika checkpoints and Moshika-RL), PersonaPlex (base and RL), NVIDIA VoiceChat-11B, MiniCPM-o 4.5, and Raon-SpeechChat. Each configuration was tested on 2,000 trials (40 topics x 10 conditions x 5 seeds).
What are the key numbers in the paper?
Direct questions raise speech onset by 0.08-0.24 over the neutral baseline and 1.5 s of inserted silence raises it up to 0.83; false facts and hazards differ from baseline by at most 0.06. When handed the floor, only 14-15% of replies to false claims challenge them and 4-7% of hazard replies give a warning.
Is the code and data available?
Yes. The paper states that stimuli, model outputs, and evaluation code are available at github.com/vocaliodmiku/take-the-floor.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are available only as captions; trial rasters and probe curves could not be inspected directly.
- Stimulus texts, the judge prompt, and exact per-cell denominators are not in the extracted text (~200 trials/cell inferred from 2,000 trials over 10 conditions).
- Several cited systems are 2026 preprints or model cards whose contents could not be verified from the text alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.