CLERC Épée v0.3: Sign Language Data Layer

CLERC Épée v0.3 is a multi-signer pilot dataset of American Sign Language (ASL): 1,200 clips from 6 Deaf signers (200 clips each). Five signers form a parallel grid of 201 phrases where renderings of the same phrase are paired by an explicit phrase_id; the sixth signer contributes 200 declarative sentences from a disjoint prompt set, adding a second register to the corpus.

 

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 gold-standard ASL gloss segmentation with temporal boundaries.

 

New in v0.3: two additional signers (6 total), corpus expanded from 600 to 1,200 clips, parallel grid widened from 150 to 201 phrases with explicit phrase_id alignment, a second linguistic register, pose-guided head-silhouette extraction (99.98% detection rate), and the cross-signer generalization benchmark re-run on all six signers with the v0.2 protocol: accuracy on a held-out signer rises from 29% to 69% (macro-F1 0.17 to 0.57) as training signers increase from 1 to 6.

 

The dataset is mirrored on Hugging Face at CLERC-DATA/epee and released under CC BY-NC-SA 4.0. Commercial licensing and full corpus access: contact florian@clerc.io.

 

Statistics: 1,200 clips, 127,545 frames, 70.86 min signed duration, 3,788 gloss tokens, 724 unique glosses.