A Python program that finds your hand, reads the fingerspelled letter, and builds words. We tested it on people it had never seen.
A random split mixes every person's video frames together, so the model sees near copies of its test questions. Testing on a person it has never seen is the honest number.
Early result from 2 people. The most common mix-ups so far: V read as U, T as S, O as C, F as B.
MediaPipe locates 21 points on your hand in each camera frame.
The points are centered on the wrist and scaled by hand size.
A scikit-learn Random Forest picks the most likely letter.
A letter counts after about half a second, then joins the word.
24 letters are covered. J and Z are left out because they need motion.