This is the official repository for the SmoothConv and DuplexConv datasets, co-developed by ASLP@NPU and QualiaLabs.
Open-source Chinese speech corpora remain dominated by single-speaker, read, or scripted data. Multi-channel, spontaneous full-duplex multi-party conversation, with overlapping speech, backchannels, interruptions, pauses, and turn transitions, is still scarce.
SmoothConv and DuplexConv are constructed from the same underlying conversational sources and are designed to complement each other. Together they provide 2,100 hours of natural Chinese dialogue in Tutoring and Social Chat. In the released directory layout, Tutoring data appear under folders prefixed with edu or Edu; Social Chat data appear under folders prefixed with none_Edu. SmoothConv (100 hours) offers high-fidelity expert human annotations for benchmarking and supervised training; DuplexConv (2,000 hours) provides large-scale LLM-assisted annotations for Speech LLM pre-training and data-driven modeling of conversational speech systems.
🎬 Demo Page · Annotation sample videos and dataset distribution charts.
- 2026/06: Official open-source release of SmoothConv (100h) and DuplexConv (2,000h).
- 2026/04/06: Released the FastTurn Test Set, constructed from a subset of SmoothConv data, for turn-state prediction evaluation.
- SmoothConv: HuggingFace
- DuplexConv: HuggingFace
- FastTurn Test Set: HuggingFace
SmoothConv contains 100 hours of naturally occurring multi-party Chinese conversations in multi-channel Tutoring and Social Chat recordings. Unlike corpora dominated by read or scripted speech, it captures realistic full-duplex dynamics: overlapping speech, backchannels, interruptions, pauses, and turn transitions. Trained experts provide fine-grained labels, including:
- Aligned transcripts with millisecond-level timestamps
- Turn-taking & overlap (speaker switches, overlaps, floor-holding)
- Pauses & sound events (laughter, coughs, breaths, silence, etc.)
- Speaker attributes (gender, age, emotion, and related interaction tags)
DuplexConv scales to 2,000 hours in the same domains, sharing a unified data design with SmoothConv. An LLM-assisted pipeline produces transcripts, speaker-aware conversational structures, turn-level interaction information, and scene-level contextual labels for large-scale Speech LLM pre-training and data-driven modeling.
Together, SmoothConv and DuplexConv bridge fine-grained human annotation and large-scale training in realistic full-duplex settings. See the Demo Page for distribution charts.
| SmoothConv | DuplexConv | |
|---|---|---|
| Annotation | Expert human | LLM-assisted |
| Duration | 100.53 h | 2,000.21 h |
| Audio files | 2,503 | 93,709 |
| Mean clip length | 144.6 s | 76.8 s |
| Clip length range | 60.0 – 634.7 s | 8.0 – 618.3 s |
| Domains | Tutoring (edu* / Edu*), Social Chat (none_Edu*) |
Tutoring (edu* / Edu*), Social Chat (none_Edu*) |
A related turn-state evaluation benchmark derived from SmoothConv, available on HuggingFace. It combines real conversational segments with 1,000 synthesized wait samples (text: DeepSeek V3; audio: IndexTTS2) to supplement the naturally rare wait class.
| Turn State | Source | Samples | Duration (h) |
|---|---|---|---|
| Complete | real-world | 14,709 | 9.64 |
| Incomplete | real-world | 3,643 | 2.15 |
| Backchannel | real-world | 3,080 | 0.42 |
| Wait | synthesized | 1,000 | 0.71 |
See the FastTurn repository for details.
For questions or collaboration: jimz@qualialabs.ai
Guidelines for the responsible collection, release, and use of SmoothConv and DuplexConv.
- Informed consent. Conversations were recorded with the knowledge and consent of participants. Personal identifiers have been removed or anonymized prior to release.
- Privacy protection. The datasets are released for academic and research purposes only. Users must not attempt to re-identify speakers or reconstruct private information from the audio or annotations.
- Intended use. SmoothConv and DuplexConv are intended for research on spoken dialogue, turn-taking, and speech understanding. They must not be used for unauthorized surveillance, impersonation, or generating deceptive content.
- Limitations & bias. Annotations may contain errors; DuplexConv labels are machine-assisted. Researchers should account for domain, demographic, and annotation bias when training or evaluating models.
- Responsible use. By using these datasets, you agree to comply with applicable laws and ethical guidelines. Report suspected misuse to jimz@qualialabs.ai.
Released under CC BY-NC 4.0.
If you use SmoothConv or DuplexConv in your research, please cite:
@article{wang2026duoconv,
title = {DuoConv: Large-Scale Chinese Full-Duplex Speech Datasets for Conversational AI},
author = {Chengyou Wang and Chunjiang He and Zhou Zhu and Lei Xie},
journal = {arXiv preprint arXiv:0000.00000},
year = {2026},
note = {Placeholder; paper forthcoming}
}