Six stages — camera to speech — each happening in under 100 milliseconds.
Try it livePosition your hand in front of the webcam. LinguaSign captures the live video feed at up to 30 FPS.
MediaPipe Hands maps 21 joints — fingertips, knuckles, wrist — producing 42 x,y coordinates per frame.
Coordinates are σ-normalised via StandardScaler so the model is position- and distance-independent.
A 3-layer MLP (256 → 128 → 37) classifies the 42 features into one of 37 Urdu letters at 98.6 % accuracy.
Accept detected letters one by one to form a word, then add words to compose a full Urdu sentence.
Pre-recorded Urdu audio reads the recognised word or full sentence aloud — instantly.