The Deaf Digital Heritage

What we are actually building, and who owns it.

What survives of a language?

Latin survived because it was written. Text carried spoken languages across the centuries, and text is why machines speak them today: every large language model was trained on the written trace of human speech.

Sign language has no written form. No books, no letters, no archive of everyday conversation. For centuries, our language and our history have lived exclusively inside our physical communities. A signed sentence existed while it was signed. Then it was gone.

Video did not fix this. Footage preserves pixels, not language. No boundaries, no glosses, no grammar. Nothing a linguist or a model can actually use.

The century that gets recorded

We are living through the first century where a language either becomes structured data or becomes invisible to the systems everyone uses.

Written languages crossed that line years ago. That is why AI writes English fluently. Sign language has not crossed it. Not because it is less rich. Because nobody built the layer.

That layer is what CLERC builds. I named it in our manifesto: the Deaf Digital Heritage, DDH. A structured, permanent corpus of the world's sign languages, built by the communities who carry them. One language at a time, one country at a time. Not archive footage. Living data: each sign segmented, labeled, validated by Deaf annotators, ready for any technology that comes next.

I say heritage on purpose. A resource is something you extract. A heritage is something you keep, maintain, and pass on. Most sign language datasets are built like resources: scraped from interpreters on YouTube, labeled far from the community, optimized for a paper deadline. We are building the other thing.

What one record actually contains

Structured is a word people use loosely, so here is what it means in practice.

Take one video of one Deaf signer telling a short story. Footage gives you a file and a duration. A DDH record gives you the sentence taken apart.

Where each sign begins and ends, in a stream that has no spaces and no punctuation. What each sign is, rather than a loose English translation of the whole clip. What the sentence is doing, not only what it says. And what the face is doing, because in sign language the face carries grammar and not just emotion.

Every one of those layers is decided by a Deaf annotator, then checked again before the record enters the corpus.

I will not detail the conventions underneath. They took years to get right, and they are the reason the corpus holds together across signers, across sessions, and across years. That consistency is the product. Anyone can film a signer. Almost nobody can hand you a corpus whose records all agree with each other.

That is the difference between a video of a language and a record of a language. The first one is a memory. The second one is something a linguist can study, a model can learn from, and a community can hand to whoever comes next.

From the community, in Deaf hands

Every video in our corpus was recorded by a native Deaf signer who chose to contribute. Every annotation was made or validated by a Deaf annotator, which is not a values statement but a technical requirement.

A heritage built by outsiders is a museum exhibit: observed, labeled from the outside, subtly wrong. Built by the community, it is the language itself, alive in a new form.

The DDH is sourced directly and exclusively from the Deaf community, and it is structurally protected: it stays in Deaf hands, and it is never up for extraction. That protection is not a wall. It is what makes the gift possible.

And the heritage stays plural. ASL is not LSF. A regional variant is not an error. Each language, each country gets its own corpus, carried by its own community. Flattening them into one convenient dataset would be efficient. It would also erase exactly what we are trying to save.

70 million, half a billion

Around 70 million Deaf people sign. Add the families, the friends, the colleagues, the teachers and the interpreters, and you reach half a billion people whose daily life touches sign language.

The DDH is built from the Deaf community. It is not built only for it.

Through controlled access, one foundation serves the whole ecosystem. Foundation model labs like OpenAI and Anthropic are missing a sign language modality: this is it. AI labs and researchers get a corpus deep enough to benchmark honestly. Entrepreneurs get a foundation to build on instead of six months of data collection before a first prototype. Universities and cultural institutions get material grounded in native signing.

It is already starting. As of August 2026, 600 of our videos are open source, our first public dataset has passed 1,800 downloads, and five hours of structured signing already power a recognition demo and a learning platform. Same corpus, several uses.

For life

A demo can be shut down. A product can pivot. The DDH has to outlive the company that started it.

So that is the standard: versioned, documented, structured to stay usable decades from now, access controlled by design. In Deaf hands, open to the world on our terms, for life.

We are at roughly 0.1% of this. One language, one country, five hours. The work is measured in decades. We plan accordingly.

The last article of this series is the hardest one. It is about what worries me. Not what AI does to jobs or to interpreters. What it could quietly do to the language itself.

Follow @CLERC to track the build.