Try the ASL fingerspelling tutor
Spell the American Sign Language alphabet to your webcam. The tutor shows you a letter, watches your hand, and tells you what to fix. It is the same model and the same feedback as the project on GitHub, running inside your browser.
- Private: everything runs on your device. The camera image is never uploaded or stored.
- Setup: good light, your signing hand raised to about shoulder height, palm facing the camera.
- First start downloads the hand tracker (about 10 MB) from Google's MediaPipe servers.
How it works
Google's MediaPipe finds 21 points on your hand in each video frame. Those points are centred on your wrist and scaled to your hand's size, then a small neural network (18,000 numbers, 160 KB) turns them into a letter. Hold the target letter steadily for just over half a second to pass. If the shape is off, it compares your fingers with how real signers make that letter and says which finger to move.
Honest limits: this browser version uses the simple per-frame model, trained mostly on one public dataset and my own hand, so it recognises some people's signing better than others'. J and Z are motion letters and are judged by their final hand shape only. Read the build log for how it was built and tested.