DuplexDrama: A Synthesized Dialogue Dataset with Scenarios, Full-Duplex Behaviors, Expressive Speech, and Sound Events
Qingxiang Guo, Wenke Fan, Shuofeng Zhao, Dawei Yang, Zhiyang Zhou, Yingxin Shang, Hongwei Cai, Zhou Wang, et al.
cs.CL
Quick verdict
Claims first 800h TTS full-duplex dialogue dataset covering personas, behaviors, emotion, sound events; solid pipeline, but no human perceptual validation.
Summary
DuplexDrama is a TTS-synthesized bilingual spoken-dialogue corpus for training full-duplex dialogue models. A four-stage pipeline expands LLM-generated persona and scenario seeds (DeepSeek-v4.1-pro), generates scripts tagged with full-duplex behaviors, seven emotion classes, and sound-event tags (Gemini-3.1-pro-preview), synthesizes speech via IndexTTS2 over a 64-voice pool, and injects environmental noise plus 52 sound-event classes at forced-alignment trigger positions under controlled SNR. Interruptions become fade-out overlaps, backchannels overlaid turns, and incompletes in-sentence pauses. The released subset holds 6,400 dialogues (800 hours; ~500 h Chinese, ~300 h English) out of over 2,000 hours produced, with 3.8% of ~362k turns carrying a full-duplex behavior. Filtering retains dialogues with speaker-consistency cosine ≥0.9; reported metrics are WER 1.8%, SpkCons 97.3%, UTMOSv2 2.57/2.52 and NISQA-MOS 3.65/3.18 (clean/mixed), with dual-LLM script scores of 4.1–4.9 out of 5. Main limitations: quality rests entirely on objective proxies and LLM judges — no human listening test, no inter-judge agreement; validator discard rates are unreported; downstream training value is claimed only through an internal citation; and the duplex behaviors are mechanically scripted rather than emergent, with the low UTMOSv2 left without a supporting ablation.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 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
- DuplexDrama is the first synthesized spoken-dialogue dataset jointly covering persona/scenario annotations, three full-duplex behaviors, persona-aligned expressive emotion labels, and script-aware sound events.
- A four-stage LLM+TTS pipeline (persona/scenario generation, tagged script generation, IndexTTS2 synthesis, FA-aligned event injection) plus a quality validator yields high-quality full-duplex training audio at scale.
- The released corpus comprises 6,400 bilingual dialogues / 800 hours, with 3.8% of ~362k turns carrying at least one full-duplex behavior.
- Objective audio metrics (WER 1.8%, SpkCons 97.3%, UTMOSv2 2.57/2.52, NISQA-MOS 3.65/3.18) and dual-LLM script scores of 4.1–4.9/5 establish dataset quality.
- The data has been validated through internal full-duplex model training (AdaptDuplex, external Zenodo citation [1]).
Strengths
Credible four-dimension novelty claim
Table 5 shows no prior corpus combines persona+scenario (①), full-duplex status (②), and sound events (③); the script-aware event tag layer is plausibly unique even granting DuplexConv's natural overlap audio.
Technically sound event/overlap realization
Sound events are injected at the FA-aligned trigger-word onset (Fig. 3) at controlled SNR on separate background channels (4 tracks per dialogue); interrupts fade out, backchannels overlay, incompletes pause (§2.3) — the right design for full-duplex training.
Internally consistent and honest reporting
Scale with a defined quality gate
Reproducible-scale release plan
Weaknesses
No human perceptual evaluation
Self-referential validation
Downstream utility unverified in-paper
Mechanical duplex semantics and unexplained mix
Reproducibility contingent and unaccounted filtering
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DuplexDrama (Zuoyebang) is a resource paper for full-duplex spoken dialogue modeling. Its central claim (Abstract, §1, Table 5) is priority: the first synthesized dialogue corpus jointly covering (i) persona+scenario annotations, (ii) three full-duplex behaviors (interruption, backchannel, incomplete), (iii) expressive speech with persona-aligned emotion labels, and (iv) script-aware sound events. Quantitative claims: 2,000+ hours produced; a released subset of 6,400 bilingual dialogues / 800 h (~500 h Chinese, ~300 h English); 3.8% of ~362k turns carrying at least one duplex behavior; quality established by objective audio metrics and dual-LLM script judging; downstream utility 'validated through internal full-duplex model training [1]' — an external Zenodo citation, not results in this paper.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 is DuplexDrama?
A TTS-synthesized bilingual (Chinese/English) spoken dialogue dataset for training full-duplex speech dialogue models. Each dialogue comes with persona and scenario metadata, tags for three full-duplex behaviors (interruption, backchannel, incomplete utterance), emotion labels, and sound-event annotations, delivered as four aligned audio tracks (two speech, two background).
How big is the DuplexDrama dataset?
The released subset contains 6,400 dialogues totaling about 800 hours (roughly 500 hours Chinese and 300 hours English), curated from over 2,000 hours produced. It has about 362,000 turns, a 64-voice pool across 13 personas, and 3.8% of turns carry at least one full-duplex behavior.
Is DuplexDrama publicly available?
At publication the authors promised to release a 6,400-dialogue subset, the audio asset bank (600 environmental clips and 2,203 sound-event clips), and their LLM-judge evaluation prompts; samples are on a demo page. Full availability should be checked on that demo page, since nothing was downloadable at review time.
How is DuplexDrama different from corpora like Fisher or CANDOR?
Fisher and CANDOR contain authentic human conversations, but without scenario or persona metadata, full-duplex behavior labels, or sound-event annotations. DuplexDrama is fully synthesized, so every dialogue ships with rich labels and separated speech/background channels — at the cost of synthetic acoustics and mechanically scripted overlap rather than natural barge-ins.
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.
- The supplied metadata lists 8 authors, while the extracted byline lists 12 (adding Weixu Wang, Lin Yang, Shuran Zhou, and Yang Song) — a metadata discrepancy, not a truncation of the paper text, which ends cleanly through reference [19].
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.