README.md

metadata

pretty_name: WebTalk-Synthetic
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - audio-to-audio
  - other
tags:
  - co-speech
  - facial-animation
  - talking-face
  - FLAME
  - synthetic
  - 3d-motion
size_categories:
  - 10K<n<100K

WebTalk-Synthetic

WebTalk-Synthetic is a dataset of synthetic in-the-wild co-speech facial motion: ~12.7 k short clips, each pairing conversational speech with a generated 3D facial-motion track in FLAME coefficient space. The facial motion is produced by an audio-driven face model from filtered in-the-wild talking audio — it is synthesized, not motion-captured. The dataset was created for and used by ViBES (CVPR 2026) to give the face expert broad in-the-wild coverage.

⚠️ Research use only. The audio is segmented from public talking-head videos. This dataset is released under CC-BY-NC-4.0 for non-commercial research only. Do not use it for commercial purposes.

Download

huggingface-cli download JuzeZhang/WebTalk-Synthetic \
    --repo-type dataset --local-dir WebTalk-Synthetic

Each modality is shipped as a single .tar (far faster to upload/sync than ~50k loose files). Extract them in place after download:

cd WebTalk-Synthetic
for f in audios audios_token_glm FLAME_coeffs_25 transcripts; do tar -xf "$f.tar"; done
# optionally: rm *.tar

Dataset structure

In the repo (as shipped):

WebTalk-Synthetic/
├── audios.tar            → audios/            (16 kHz mono WAV, one per clip)
├── audios_token_glm.tar  → audios_token_glm/  (GLM-4-Voice audio tokens, one .npy per clip)
├── FLAME_coeffs_25.tar    → FLAME_coeffs_25/   (synthetic FLAME face motion, one .npz, 25 fps)
├── transcripts.tar        → transcripts/       (one .txt per clip)
├── train.txt              (12,075 clip stems)
├── val.txt                (290 clip stems)
├── test.txt               (304 clip stems)
├── README.md
└── LICENSE

audios_token_glm/<stem>.npy is a (N,) int64 array of GLM-4-Voice discrete audio tokens, provided so you can skip re-running audio tokenization.

Every modality is keyed by a clip stem of the form <session>_<segment> (e.g. 202008647_0001); audios/<stem>.wav, FLAME_coeffs_25/<stem>.npz, and transcripts/<stem>.txt all refer to the same clip.

• 12,669 clips total (12,075 train / 290 val / 304 test).

FLAME motion format (FLAME_coeffs_25/<stem>.npz)

Key	Shape	dtype	Description
exp	(T, 100)	float32	FLAME expression coefficients
shape	(T, 100)	float64	FLAME shape coefficients
pose	(T, 6)	float32	head pose (3) + jaw pose (3), axis-angle
mocap_frame_rate	scalar	int64	25

T is the per-clip frame count at 25 fps.

Audio

16 kHz mono PCM WAV, ~8 s per clip.

Usage

Load a clip directly:

import numpy as np, soundfile as sf

stem = "202008647_0001"
audio, sr = sf.read(f"audios/{stem}.wav")          # 16 kHz mono
coef = np.load(f"FLAME_coeffs_25/{stem}.npz")      # exp / shape / pose
text = open(f"transcripts/{stem}.txt").read()

The full ViBES preprocessing recipe (audio tokenization, face VQ-VAE tokenization, and building the training-ready HuggingFace dataset) is documented in docs/1-data/webtalk_synthetic.md.

Intended use & limitations

• Intended for non-commercial research on co-speech facial animation, audio-driven face generation, and conversational virtual humans.
• The facial motion is model-generated, not ground-truth capture; it reflects the biases and failure modes of the audio-driven face model that produced it.
• Audio originates from public talking-head videos; treat it accordingly and do not attempt to re-identify speakers.

License

CC-BY-NC-4.0 — non-commercial research use only.

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},
}