qualialabsai/smoothconv-duplexconv

SmoothConv and DuplexConv are two Chinese speech datasets designed to help researchers build and test AI systems for real spoken conversations. Most existing Chinese speech datasets contain single speakers reading scripted text, these two cover genuine multi-party dialogue where people talk over each other, interrupt, laugh, and hand off the floor in natural ways. That kind of data is scarce, and it is necessary for training AI that can hold realistic two-way conversations. The two datasets come from the same original recordings and cover two scenarios: tutoring sessions and casual social chat. SmoothConv is the smaller, higher-quality part: 100 hours of audio with labels created by trained human annotators. Those labels include timestamped transcripts accurate to the millisecond, notes on who overlapped whom and when, markers for pauses, laughter, coughs, and breathing, and per-speaker tags for gender, age, and emotion. DuplexConv is the larger part: 2,000 hours of audio from the same domains, with labels generated by an AI-assisted pipeline instead of humans. The AI-assisted labels include transcripts, speaker-aware conversation structure, and scene-level context information. The two datasets together are meant to complement each other: SmoothConv for benchmarking and supervised training where label quality matters most, DuplexConv for large-scale pre-training where volume matters. A related resource called the FastTurn Test Set is also provided. It is a benchmark for evaluating whether a model can predict what kind of turn a speaker is in the middle of: a complete utterance, an incomplete one, a short backchannel response, or a wait signal. It contains about 22,000 real clips from SmoothConv plus 1,000 synthesized wait examples added to balance the classes. All three resources are available on HuggingFace. The datasets are released under Creative Commons Attribution NonCommercial 4.0, which means they can be used for academic research but not for commercial products. The project is a collaboration between the ASLP lab at Northwestern Polytechnical University and QualiaLabs.