Four Months, Less Than $1,000, and One Conviction: Data Comes First

In four months, with less than $1,000, CLERC built a browser demo that recognizes 15 ASL signs. The first benchmark: 84% corpus accuracy, 0.86 Macro-F1, 63% hold-out accuracy, and up to 71% on a signer the model had never seen. You can try it at clerc.io/demo.

That last number matters most. Unseen signers are where most sign language recognition demos break. Lighting, rhythm, handshape, body geometry, facial grammar, regional habits: the model either learned the language signal or it learned the people in the training set.

One clarification before the demo: CLERC is not becoming a recognition product company. The demo exists to prove the data layer. If the data is diverse, Deaf-built, carefully annotated, and validated, even a small model can start to generalize.

What we built

Less than a thousand dollars bought us no institutional cushion. What moved the work forward was a group of Deaf signers recording real ASL, and a small team organizing, annotating, and validating thousands of videos. The public datasets and this demo came from the same pipeline.

These numbers are not a victory lap. They are a baseline we can publish without hiding the limits:

• Corpus: 84% accuracy - Macro-F1: 0.86
• Hold-out: 63% accuracy
• Unseen signer: up to 71% accuracy

Small vocabulary. Early model. Honest benchmark. That is more useful than a polished demo with no test discipline.

Why build a recognition demo at all?

Data has a visibility problem. A dataset looks like spreadsheets, video files, and annotation logs, rows of careful work whose value almost no one can see by looking at it. A working application changes that in a second.

When a sign is recognized correctly, the visible event is the prediction. The important thing is the evidence beneath it: clean videos, consistent labels, signer diversity, validation decisions, versioned data. The demo makes that invisible work inspectable.

The real problem

Sign language has not had its ImageNet moment. Text models advanced when enough structured text existed to train them. Image models advanced when the images did. Sign language still lacks a corpus deep enough, diverse enough, and structured enough for modern AI.

The gap is not goodwill, and it is not an accessibility feature added at the end. It is a training data gap. A model cannot learn grammar that was never captured, and it cannot generalize to Deaf people it never saw represented in the data.

Deaf people have to build Deaf AI

None of this exists without the Deaf community. Native signers recorded the videos, validated the language, and caught errors a hearing team would miss: gloss boundaries, regional variation, facial grammar, register, rhythm, intent.

I am Deaf, and I do not treat that as a credential. I treat it as a requirement. Whoever builds the data layer writes the grammar of every model trained on top of it, so sign language AI has to be built with Deaf people, not merely for them.

What this unlocks

Recognition is the first proof, not the destination. The same corpus that reads fifteen signs today can support translation engines, educational tools, searchable sign libraries, digital assistants, accessible customer service, and the multimodal systems arriving next.

We will keep expanding the corpus, widening signer diversity, deepening the annotations, and opening it to researchers and developers. The benchmark will move because the data will move.