THE CLERC DEMO: From Our Database to Your Use Case
Sign language has never had an ImageNet moment. No standardized corpus, no linguistic ground truth, no data layer solid enough to train on at scale. That is not just a gap in the field. It is a structural absence, and it is the reason every sign language AI product built today is fundamentally fragile.
CLERC is building the foundation that does not exist yet.
What the demo shows
We ran 100 videos with three native Deaf signers, producing 1,100 annotated glosses and achieving 90% recognition accuracy. The model performs because the data is clean. That is the core thesis of everything we are building: intelligence follows data quality.
The demo is not a proof of concept. It is a demonstration that the pipeline works end to end — from recording through annotation, validation, and training, to a model that generalizes beyond the signers it was trained on.
How the pipeline works
Most sign language AI demos are built on shortcut data. Researchers record a small number of signers in controlled lab conditions, label the clips with approximate English glosses, and train a model that learns to pattern-match on those specific signers in those specific conditions. The model performs well in the demo. It collapses when a new signer appears, or the lighting changes, or the vocabulary goes out of distribution.
CLERC works differently because the data is built to a different standard.
Recording: All signers in the CLERC corpus are native Deaf signers — people for whom ASL is a first language, not a learned skill. Native signers produce natural, unconstricted signing with full non-manual grammar. Most existing datasets record Deaf signers signing slowly and carefully for a camera. CLERC captures natural register across a range of domains and topics.
Annotation: The annotation pipeline captures all five parameters of ASL sign production: handshape, location, movement, palm orientation, and non-manual markers. This is not a video-to-English translation. It is a linguistically grounded transcription of what is actually being signed, including the facial grammar that carries clause-level information. Annotations are produced by trained annotators under the supervision of ASL linguists.
Validation: Every annotation goes through a multi-stage validation process before entering the corpus. Human validators — all of whom are fluent in ASL — verify the annotation against the video. Disagreements trigger review. The goal is not speed but correctness, because errors compound silently in training data.
Versioning: The corpus is versioned from day one. Every release is tagged, documented, and reproducible. Researchers and AI teams using CLERC data can point to an exact version of the corpus that produced their results — a requirement for scientific reproducibility that most AI training datasets do not meet.
What the results actually mean
90% recognition accuracy is meaningful in context. The field benchmark for open-vocabulary continuous sign language recognition — meaning recognition of unrestricted signing from signers the model has not seen before — is significantly lower. Most published results operate on constrained vocabularies or test on signers included in training data.
The CLERC demo model was trained on three signers and tested on signing it had not seen from those signers before. It was not tested on held-out clips from the same sessions. The generalization matters.
The 1,100 glosses produced from 100 videos gives a sense of the annotation density. That is approximately 11 signs per video on average, which reflects natural conversational signing rather than isolated sign elicitation. Continuous, connected signing is harder to annotate and harder to train on — and it is the only kind of data that produces models that work in real conditions.
What you can build on this foundation
The CLERC corpus is infrastructure. It is not a product itself. What you can build on top of it depends on your use case.
Translation tools: A corpus with linguistically grounded annotations is the training foundation for ASL-to-English translation models that go beyond pattern matching on vocabulary and handle the spatial grammar of connected discourse.
Accessibility products: Real-time sign language recognition for communication tools, automated captioning for signed content, interfaces that respond to sign language input rather than requiring spoken or written language.
Research systems: Pose estimation benchmarks grounded in linguistically validated annotations, models for non-manual marker detection and classification, sign spotting systems for video retrieval — each of these requires the kind of structured, validated data that CLERC produces.
Foundation model integration: The modality gap in current foundation models is well understood by the labs building them. What they are missing is data infrastructure that meets the standards required for foundation model training. CLERC is designed explicitly for this use case, with provenance tracking, consent compliance, and versioned releases.
The database is what makes any of it viable. We provide the foundation. You build the experience.
Where we go from here
The EPEE v01 release is available on Hugging Face. It is the first public release from the CLERC corpus — a working demonstration that the pipeline produces the kind of data we claim it does.
The roadmap beyond EPEE includes expanded signer diversity, broader domain coverage, and deeper annotation layers for spatial grammar and discourse structure. Each release is designed to be a production-quality addition to the corpus, not an exploratory sample.
If you are working on sign language AI — whether you are an AI lab, a research team, a translation product, or a developer building accessible communication tools — the data problem is the first problem you will hit. We have spent years building the infrastructure to solve it.