I'm a Deaf founder building sign language data infrastructure. After ten years in sign language tech (signing avatars for Airbus, the French government, and the 2025 Deaflympics), I became convinced the bottleneck isn't models, it's data. Almost every existing corpus is research-only, interpreter-based, or unstructured video.
EPEE is our open benchmark subset: 600 ASL clips from 4 native Deaf signers, with sign-level segmentation, gloss labels, 128 keypoints per frame, and 150 parallel phrases. It ships with a cross-signer benchmark: recognition on an unseen signer climbs from 22% to 59% as you add training signers.
I'm a Deaf founder building sign language data infrastructure. After ten years in sign language tech (signing avatars for Airbus, the French government, and the 2025 Deaflympics), I became convinced the bottleneck isn't models, it's data. Almost every existing corpus is research-only, interpreter-based, or unstructured video.
EPEE is our open benchmark subset: 600 ASL clips from 4 native Deaf signers, with sign-level segmentation, gloss labels, 128 keypoints per frame, and 150 parallel phrases. It ships with a cross-signer benchmark: recognition on an unseen signer climbs from 22% to 59% as you add training signers.
Try it on Hugging Face: https://huggingface.co/datasets/CLERC-DATA/epee
The dataset is also archived on Zenodo with a citable DOI. I wrote up why most sign language datasets can't be used commercially here: https://clerc.io/blog/the-sign-language-ai-dataset-landscape
Happy to answer anything about ASL linguistics, annotation tooling, or building datasets with the Deaf community.