The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Motion-Omni

Training and evaluation data for Motion-Omni: End-to-End Joint Speech and Full-Body Motion for Spoken Dialogue.

Chengqian Ma1 · Wei Tao2 · Haoyu Zhang3 · Yiwen Guo4,*

1Peking University · 2LIGHTSPEED · 3The Chinese University of Hong Kong, Shenzhen · 4Independent Researcher *Corresponding author

Code: github.com/step-out/Motion-Omni


Status

We are preparing the release and expect to publish it by December 2026.

What will be released

Training pairs — 422,856 samples, 1,402 hours. Each pair holds a spoken instruction response in a consistent target voice together with co-speech motion: facial expression plus hand, upper-body and lower-body movement. The motion is produced by running a reusable motion teacher over every response waveform of a speech-instruction corpus, then scored with a dual-metric quality measure that also induces the training curriculum. The quality ranking will be included, since the curriculum cannot be reproduced without it.

Motion is stored as SMPL-X body parameters with FLAME facial expression coefficients.

SwDA-500. The evaluation prompt set: 500 open-ended conversational prompts drawn from an external corpus, so the evaluation is not scored on prompts the model was tuned against.

Downloads last month
31