DuplexConv
2,000-hour Chinese multi-channel full-duplex conversational speech dataset with LLM-assisted annotations
Duration
2000.21 hours
Languages
1
Sample Rate
48 kHz
Published
2026-06
Description
1Contains 2,000.21 hours of naturally occurring multi-party Chinese conversation recordings across tutoring and social chat scenarios
2Recorded in multi-channel environments and captures realistic full-duplex conversational behaviors such as overlapping speech, backchannels, interruptions, pauses, and dynamic turn transitions
3Audio is provided as dual-channel WAV; a sampled file was 48 kHz, 16-bit PCM
4Includes 93,709 audio files, with a mean duration of 76.84 seconds and an utterance duration range from 8.0 to 618.3 seconds
5Each audio file is paired with JSON annotations covering transcripts, speaker-aware conversational structures, turn-level interaction information, and scene-level contextual labels
6Uses an LLM-assisted annotation pipeline and targets Speech LLM pre-training, conversational speech understanding, turn-taking modeling, full-duplex spoken dialogue systems, and multi-party conversational AI research
7Released under the CC BY-NC 4.0 license
Language Details
| Language | Duration |
|---|---|
| Chinese | 2000.21 hours |
Publisher
License & Commercial Use
License
CC BY-NC 4.0
Commercial Use
Non-commercial only