YouTube_Talking

YouTube_Talking is a large in-the-wild conversational dataset: 2,983 English single-speaker talking-head videos (interviews + speeches) with derived multimodal annotations — speech audio, FLAME face motion, SMPL-X body, HaMeR hands, active-speaker segments, and motion tokens. It is one of the datasets used by ViBES (CVPR 2026).

⚠️ Research use only. The source videos are copyrighted YouTube content and are NOT redistributed — you download them yourself from the URL list. We do ship the expensive model-output annotations (TalkNet active-speaker, 4D-Humans / MHR body, FLAME face, fitted SMPL-X, HaMeR hands) so you can skip hundreds of GPU-hours of re-processing. Cheap/deterministic steps (audio extraction, tokenization, HF packing) are reproduced with the shipped scripts. License: CC-BY-NC-4.0.

Contents

YouTube_Talking/
├── video_urls.csv             (2,983 rows: id,url,language,type,body_parts,num_people,duration_min)
├── download_youtube_talking.py (downloader for the source videos)
├── talknet_pywork.tar         (TalkNet active-speaker results, pywork/ per video)
├── 4d_humans_results.tar      (4D-Humans body recovery, per-video .pkl)
├── mhr_results.tar            (MHR body, per-frame .npz)
├── smplxflame_25.tar          (fitted SMPL-X body+face @25fps)
├── FLAME_coeffs_25.tar        (FLAME face coefficients @25fps)
├── hamer_results.tar          (HaMeR hand recovery, per-frame .pkl)
├── train_processed.txt        (346 fully-processed clip ids)
├── val_processed.txt / test_processed.txt   (9 / 11)
├── train_unprocessed.txt / val_unprocessed.txt / test_unprocessed.txt
├── README.md
└── LICENSE

Each *.tar extracts to its directory in place (tar -xf <name>.tar). The talknet_pywork.tar gives talknet_output/<id>/pywork/{scores,tracks,faces,scene}.pckl. The *_processed.txt splits list the clips with full multimodal processing.

Not shipped (regenerate locally): the videos themselves, the extracted audio (audios/, audios_original/), speaking-segment crops, transcripts, and the motion/audio tokens — all cheap to recompute from the downloaded videos + shipped annotations using the scripts below.

How to use

1. Download videos — python download_youtube_talking.py --url_csv video_urls.csv --output_dir videos
2. Extract the shipped annotations — for f in *.tar; do tar -xf "$f"; done
3. Reconstruct audio / tokens / HF dataset from the videos + annotations — follow the step-by-step recipe in the ViBES repo: docs/1-data/youtube_talking.md.

All preprocessing scripts live in the ViBES repo under preprocess/.

License

CC-BY-NC-4.0 — non-commercial research use only. Source videos remain under their original YouTube terms; this release redistributes none of them.

Citation

@inproceedings{zhang2026vibes,
  title={ViBES: A Conversational Agent with Behaviorally-Intelligent 3D Virtual Body},
  author={Juze Zhang and Changan Chen and Xin Chen and Heng Yu and Tiange Xiang and Ali Sartaz Khan and Shrinidhi Kowshika Lakshmikanth and Ehsan Adeli},
  booktitle={CVPR},
  year={2026},
}