BEAT Audio-to-Blendshape (log-mel / ARKit 51)

BEAT Audio-to-Blendshape Paired 80-bin log-mel speech features and 51-dim ARKit blendshape tracks, derived from BEAT v1 (english v0.2.1). 6,357 speech segments / 55.5 hours / 30 speakers. This is the exact data our audio-driven face model was trained on, in the exact train/validation partition it used — not a re-cut sample. Contents Segments 6,357 (train 5,970 · validation 387) Duration 55.5 h (train 52.5 h · validation 3.0 h) Speakers 30… See the full description on the dataset page: huggingface.co/datasets/prateek-0-gupta/beat-audio2blendshape.