Teaching machines to sign.
Since 2012.
We built our first neural network for sign language in 2012 — long before AI was fashionable. Today we pursue a bigger vision: a hub connecting sign language with communication, business, science and technology — one place where interpreters, researchers, Deaf organizations and companies meet.
At migam.ai we develop automatic translation into sign languages: text becomes an utterance of a photorealistic 3D avatar, with Deaf native signers and linguists overseeing the quality of every stage.
What our model stands on
Years of Migam interpreters' work are a unique foundation: hundreds of hours of recordings, our own studio and a Polish Sign Language corpus built over a decade. The model grows within the NVIDIA Inception program, and we validated the methodology in a paid pilot with a global technology company.
Every country signs differently. We have grammatical descriptions of nine sign languages — from Polish to American — and an architecture designed from day one for many languages, not just one.
The project is supported by a scientific board of sign language researchers from Europe, the USA and Israel, and the language team is led by a Chief Linguist educated at the world's leading Deaf university.
Evolution of the avatar
From a point cloud to a photorealistic character — this is what every sign language utterance really goes through. Real screenshots from our pipeline, bloopers included.
Representation, not "text plus gestures"
Sign languages are full visual-spatial languages — you can't just "bolt them onto" text. That's why we build a representation layer: from grammar, through motion tokens, to photorealistic rendering.
One sign = 5 simultaneous channels
That's why captions or "waving hands" aren't enough — the model must control all of them at once.
Models change. The representation stays.
The architecture is independent of any single AI model — swappable parts can be replaced as technology advances, while our core keeps working.
Not just a product. We're learning to understand languages and communication better.
We document every stage like a research paper — with assumptions, literature and measurable quality thresholds. That knowledge stays: grammar descriptions, sense lexicons and a methodology the next sign languages can build on.
How does the avatar know which meaning of a sign to pick?
One sign can carry many senses — like the word "bank" in English. Word Sense Disambiguation (WSD) is choosing the right meaning in context. Three layers handle it: the Lexicon (which senses exist — a linguistic decision), the Corpus (how they are used — a corpus decision) and WSD rules (how to choose from the source text — an engineering decision).
Two language tracks, one workshop
Join us
Do you research sign languages? Run a corpus, a lexicon or sign linguistics courses? Join the workshop — joint publications, grant applications and a methodology ready to adapt to your language.
kontakt@migam.org →Since when and with whom
Strong partners and hard evidence — no overpromising.
These aren't plans. We sign letters of intent.
The hub is already growing: we have signed letters of intent with partners from five European countries — from the Benelux to the Balkans — who want to develop their sign languages with us. The collaboration model is simple: we bring the technology, the AI pipeline and a proven methodology — the partner brings linguists, Deaf native signers and knowledge of their market. Together we walk the path from recordings and annotation to working translation. We collaborate with Deaf organizations, companies and universities — including joint grant applications.