SignBridge is a real-time sign language translator: your browser webcam feeds MediaPipe’s hand landmarker, which turns each frame into 130 normalized keypoints, and a rolling buffer of 75 frames is downsampled to 30 and sent to a FastAPI server where an LSTM predicts the sign. The frontend only displays a word when confidence is above 0.8, the label isn’t “idle”, and there’s enough hand motion, while the ml/ folder covers dataset extraction, custom recording, training, and live testing, with
utils.py keeping the keypoint math consistent between Python and JavaScript.