Founder & CEO of CLERC - the data infrastructure layer for sign language in multimodal AI.

Text and images have a data layer. Sign language does not. I build the one that is missing: native-signer video corpora, annotated to a versioned schema, released open for research and licensed to foundation model labs.

Native LSF signer. ~10 years in sign language technology. I write the pipeline myself.

→ huggingface.co/datasets/CLERC-DATA/epee

An open, AI-grade ASL keypoint corpus. Free to download, free for research.

• 1,200 clips · 6 Deaf signers · 71 minutes of signing - 200 clips per signer
• A 201-phrase parallel grid across five of the six signers: the same sentence, signed by each, so signer variation can be isolated from phrase variation. The sixth contributes 200 declarative sentences on a separate prompt set, a second register
• 127,545 frames, 3,788 gloss tokens, 724 unique glosses
• (n_frames, 128, 3) keypoint tensors - MediaPipe Hands, Pose, Face and FACE_OVAL silhouette
• Privacy-preserving by design - eyes, mouth and head outline only, no internal facial features
• Gold-standard gloss segmentation with temporal boundaries
• Ships with a cross-signer generalization benchmark
• CC BY-NC-SA 4.0 · commercial licensing at florian@clerc.io

Sign language AI has been held back by datasets that are small, single-signer, or locked behind institutional agreements. Épée is the piece I wanted to exist and did not: openly downloadable, schema-validated, and built with Deaf signers who are paid and credited.

Recognition accuracy on a held-out, never-seen signer, as a function of how many signers are in the training set:

A signer the model has seen scores 73%, so the stranger gap closes from 43 points at one training signer to 3 points at six, and the curve has not flattened.

Signer diversity, not clip count, is what moves cross-signer generalization. Sign language AI is data-bound, not model-bound. That is the whole thesis behind CLERC, and the benchmark to reproduce it ships inside the dataset.

Corpus production - I design and run the annotation and QA pipeline end to end: gloss segmentation with temporal boundaries, a versioned schema, validation that refuses non-conformant data at write time, review queues and quality tiering. Python throughout.

Pose & keypoints - MediaPipe Hands/Pose/Face, 128×3 per-frame keypoint tensors, privacy-preserving face representation, SMPL-X / MANO body and hand modelling, NumPy/SciPy signal processing over pose streams.

Models & evaluation - isolated sign recognition baselines, automatic sign-boundary segmentation, cross-signer generalization benchmarks, GPU data-augmentation experiments, per-gloss precision/recall analysis.

Release engineering - versioned public dataset releases, reproducible benchmarks, and open tooling so anyone can load the corpus in five minutes.

👉 epee-asl-toolkit - the open Python toolkit for the dataset: loaders, keypoint visualization, inter-signer variability analysis.

Built Keia, a signing avatar deployed at Airbus, Thales, Macif, TotalEnergies and for the French government. Co-founded Spokhand (~15M views). HEC Paris and UC Berkeley.

clerc.io · florian@clerc.io · Hugging Face · LinkedIn · X @FlorianMeloux

Training a multimodal model and sign language is a gap? Email me.

_{Keywords: sign language dataset · American Sign Language (ASL) · Langue des Signes Française (LSF) · sign language recognition (SLR) · sign language translation (SLT) · gloss annotation · open dataset · Hugging Face · MediaPipe keypoints · pose estimation · multimodal AI training data · Deaf AI accessibility · AI data infrastructure}