About signAR

Speech recognition works because there are millions of hours of transcribed audio. Sign language recognition mostly does not work, for one reason: the equivalent data does not exist. Pakistan Sign Language in particular has almost no public, machine-readable dataset at all.

signAR is an attempt to change that from the ground up. Anyone can record a sign in their browser. Each clip is reviewed by a human, added to a public dictionary, and released as part of an open dataset. The same clips train the recognition model, so the tools get better in step with the collection.

Built with the community, not about it

Sign languages belong to Deaf communities. A hearing-led dataset built by scraping videos without consent would be both worse technically and wrong ethically. Every recording here is contributed deliberately, with explicit consent, by someone who chose to give it. Contributors keep ownership of their recordings and can ask for them to be removed.

Where it is today

0 signs have at least one clip, from 0 contributors, totalling 0 recordings. That is early. A sign needs roughly twenty good clips from different people before a model recognizes it reliably, which is why the site keeps pointing you at the signs that are short.

What is honestly hard

Word-by-word playback is not translation. Sign languages have their own grammar, spatial reference and non-manual markers that no per-word clip sequence captures. The translator here is a vocabulary and lookup tool, and we would rather say that plainly than oversell it.