ASL Live

Karen's project. Sign to your camera and the computer tells you what it saw. Practice your first 250 signs in American Sign Language, and fingerspelling too, with instant feedback. It runs in your browser, and no video ever leaves your device.

How it works

Your browser tracks 543 points on your hands, face and body in every video frame. When you lower your hands, or hold the last handshape still, a recognition model reads the movement and ranks the 250 signs it knows.

Fingerspelling works the same way: spell a whole word letter by letter, lower your hands, and a second model reads out the letters it saw. That model was trained for ASL Live on about 100,000 fingerspelled phrases and reads held-out phrases with 79% character accuracy.

The 250 signs are everyday first words such as mom, milk, happy, dog and thank you. They come from PopSign, a project built to help hearing parents learn to sign with their Deaf children.

Your privacy

The camera feed is processed on your own device. Nothing is recorded, uploaded or stored on a server. Learn keeps a list of the signs you have practiced in this browser only, and you can clear it at any time.

What it can and cannot do

  • It recognizes single signs from a fixed vocabulary of 250, and fingerspelled words one at a time. It does not yet follow continuous signing.
  • Fingerspelling is read letter by letter and can drop or swap a letter, especially in long words, so check names and numbers it reads.
  • ASL carries meaning in facial expression and body movement as well as the hands. This tool reads one sign at a time and misses much of that grammar.
  • It is a learning aid and a demonstration. It is no substitute for a qualified interpreter or for learning from Deaf signers.
  • Good light, a plain background and your upper body in view all help.

Feedback from Deaf and hard of hearing signers is especially welcome at [email protected].

Credits

Recognition model: the first-place solution (by hoyso48) to Google's Isolated Sign Language Recognition competition on Kaggle, as published on Hugging Face under the MIT license, converted here to run in the browser.

Training data: PopSign ASL v1.0, from Georgia Tech, the National Technical Institute for the Deaf at RIT, the Deaf Professional Arts Network and Google, released under CC BY 4.0.

Fingerspelling model: trained by ThinkMorphic for ASL Live on Google's ASL Fingerspelling dataset (Google and the Deaf Professional Arts Network, CC BY 4.0). It scores 0.79 normalized character accuracy on held-out phrases, close to the 0.80 of the top competition entries.

Hand, face and body tracking: Google MediaPipe (Apache 2.0). Model runtime: ONNX Runtime Web (MIT).

ASL Live was Karen's idea, built by Karen and her dad, Joe Fernandez. A ThinkMorphic project, ThinkMorphic.com.