Why Data, Not the Model
I've spent ten years in the Sign Language AI space. I've sat in rooms with researchers, AI engineers, linguists, and technologists from across the field, people who've dedicated their careers to making machines understand sign language. Every single one of them, at some point, said the same thing: it's not a technology problem. It's a data problem.
And yet the industry keeps chasing better models.
The five-year-old analogy
Think of AI like a five-year-old child you're trying to teach sign language. You can give that child the best brain in the world, the most sophisticated learning capacity, perfect motor skills. But if no one sits down and actually teaches them the signs, if no one shows them what each gesture means, in context, with variation, with the spatial grammar and facial expressions that make sign language what it is, that child will learn nothing.
The lesson is the data. Without it, the intelligence is irrelevant. This is where every major Sign Language AI project breaks down: not in the architecture, not in the compute, but in the lesson itself.
What "data" actually means here
When I say data, I don't mean raw video footage of someone signing. That's the equivalent of recording a lecture in a language you don't speak and hoping the recording somehow learns from it. It doesn't, and the AI doesn't either.
What the AI needs is structured, labeled, high-quality data. Each sign segmented, each gloss identified, the spatial relationships mapped, the facial expressions documented (because they carry grammatical meaning in sign language, not just emotion), the regional variation tracked, the register captured.
This is not a small task. It is, in fact, the entire task.
Sign language is not a simplified version of spoken language. It is a complete, rich, complex linguistic system that uses three-dimensional space as grammar, that combines handshape, movement, location, orientation, and non-manual markers simultaneously. A single sign can carry what would require an entire clause in English to express, and no current corpus captures that depth at scale.
There is no shortcut through this complexity. You can't approximate your way to understanding a language this rich. You have to build the data layer honestly, from the ground up, with people who actually know the language.
What I learned building the first generation
Before CLERC, I was part of the team that launched Keia, a platform designed to make websites accessible in sign language through a 3D signing avatar. We deployed it for major organizations: Airbus, Thales, Francaise des Jeux, TotalEnergies, and the French government. The avatar worked, it signed, and it was technically functional.
But it only worked within very narrow, controlled domains: government announcements, energy sector vocabulary, specific and bounded scripted content. The moment you tried to scale beyond those predefined topics, it fell apart. Not because the avatar was bad or the model was weak, but because the data underneath it was too thin, too narrow, too carefully controlled to reflect how Deaf people actually sign in the real world.
That experience confirmed what every researcher had already told me: the bottleneck isn't the output, it's what you feed in. And it's the reason I built CLERC around data, not around a product.
Building from the Deaf community, for everyone
Collecting sign language data is difficult, and finding the right people to produce it is even harder.
In France, roughly 60 to 80% of the Deaf community faces functional difficulties with written French, the dominant language used in most technology interfaces and data collection workflows. That creates a structural barrier that simply doesn't exist in spoken language data collection. In the United States, the ASL community represents one of the most valuable and well-documented sign language ecosystems in the world, which is one reason CLERC started there.
But beyond geography, there's a deeper principle that drives everything we do at CLERC.
Sign language is a living language. It varies by region, by generation, by context, by community. It isn't a standardized code you can document once and call done. It shifts, it evolves, and it carries culture in ways that no external observer can fully capture. The only way to build a data layer that reflects this reality is to build it with the Deaf community, not just about them.
This is why CLERC is Deaf-led, and it's not a value statement or a diversity initiative. It's a technical requirement. When a hearing team works on labeling sign language, they're making thousands of judgment calls they're not equipped to make: where does one sign end and another begin? Is this a regional variant or an error? Is this facial movement grammatical or expressive? These aren't edge cases, they're the core of the problem, and getting them wrong compounds silently until the entire dataset is compromised.
We are Deaf. Our labeling team is Deaf. Our signers are native. That's not a talking point, that's the reason our data is honest. And you can see what that looks like in practice: our recent demo shows how raw video, captured by native Deaf signers, becomes structured, labeled, high-quality data that an AI can actually learn from. That's the path from raw data to real understanding, and it starts with the community.
Why data, not the model, and what it changes
CLERC is not building a translation app. We're not competing with accessibility companies or research labs. We're not in the model business at all.
We're building the data infrastructure that makes Sign Language AI technically possible, for anyone who wants to build on top of it: translation tools, 3D avatars, SLR benchmarks, lesson generators, educational platforms, research corpora. The use cases are not ours to limit; they belong to the builders, the researchers, and the Deaf communities who will shape them.
Our job is to build the layer underneath: clean, structured, native-reviewed, scalable data. The kind that bridges the Deaf world and the hearing world, not by simplifying sign language, but by capturing it in all its richness.
Sign language has never had its ImageNet moment. No standardized corpus, no linguistic ground truth, no data layer solid enough to train on. We're building that moment, and we're doing it the only way that works: from the inside.
The model is not the problem. The data always was, and we're fixing it.
Follow @CLERC to track the build.