clercÉpée v0.4

Read the Hands.

Open dataÉpée v0.4 on Hugging Face +
Open model86,000 parameters
=
Sign language recognitionin your browser

That is all it takes. Anyone can build this. This model learned only from Épée, 2,400 clips by 8 Deaf signers, and has never seen you. How many signs can it read from you in 60 seconds?

Webcam · runs in your browser

The AI reads you right here, in this tab. The video never leaves your device.

Upper body in frame, hands visible, good light.

One tap here
A quick human check, once per visit, so your minutes can be saved as dots.

60 seconds
A word appears: sign it, hands down after. The AI reads you, the next word comes. No sign is shown up front.
Turn the camera on first.

What this is

An example of what the open Épée v0.4 release already allows: a small recognizer trained on its keypoints alone, that reads people it has never seen. Épée is the open slice of the CLERC corpus, free for research under CC BY-NC-SA 4.0.

The model

The BiLSTM of the Épée benchmark: hands and arms only, 24 signs, about 86,000 parameters. Here it reads 12 of them, one word at a time. Each sign is read twice, as filmed and mirrored, so it does not matter which hand leads.

Your data

The video never leaves your device. Pressing Start saves the session's keypoints as it happens: the dots of your hands, body and face, encrypted on arrival. No image of you is ever stored, shared or sold.