Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23587708Rights:
No information provided; contact the repository owner
CLERC Épée v0.4 is a multi-signer pilot dataset of American Sign Language (ASL): 2,400 clips from 8 Deaf signers (300 clips each). Five signers form a parallel grid of 302 phrases where renderings of the same phrase are paired by an explicit phrase_id; three signers contribute 300 sentences each from their own prompt sets, disjoint from the grid and from each other. Each clip is represented as a (n_frames, 128, 3) keypoint array combining MediaPipe Hands, Pose, Face (eyes + mouth), and FACE_OVAL head silhouette, with privacy-preserving design (no internal facial features beyond eyes and mouth). Annotations include ASL gloss segmentation with temporal boundaries. Correction in v0.4: v0.1 to v0.3 shipped with crossed hand slots (61.7% of the decidable v0.3 clips had the two hands swapped). In v0.4 every clip is anchored to the pose wrists and the built files are checked independently: 0 crossed out of 2,395 decidable clips, recorded in a new hand_slots column. Users who trained on earlier versions should re-download. New in v0.4: two additional Deaf signers (8 total), corpus doubled from 1,200 to 2,400 clips, parallel grid widened from 201 to 302 phrases, three own registers, and the cross-signer benchmark re-run with hand slots fixed in train and test. On a held-out signer, accuracy rises from 42% to 81% (macro-F1 0.28 to 0.74) as training signers increase from 1 to 8. The effect of the correction is reported separately: on the same held-out signer, the v0.3 recordings score 69.4% as published, 75.7% with hand slots fixed, and the full v0.4 pool reaches 81.4%. These numbers are not comparable to the v0.3 record. One v0.3 clip was withdrawn for privacy (a fill-in prompt answered with personal information) and is not part of this version. The dataset is mirrored on Hugging Face at CLERC-DATA/epee (tag v0.4) and released under CC BY-NC-SA 4.0. Commercial licensing and full corpus access: contact florian@clerc.io. Statistics: 2,400 clips, 323,872 frames, 179.93 min signed duration, 10,441 gloss tokens, 1,501 unique glosses.
Keyword(s):
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)