Allow camera access and sign one of the words below. Every frame is reduced to a MediaPipe skeleton before it reaches the model, so a bidirectional LSTM trained on the WLASL dataset recognizes you — no video ever leaves your device.
Word-level recognition across 18 signs, ~8× the random baseline — a deliberately scoped proof of concept, not a finished translator, so expect misses. Same MediaPipe→LSTM pipeline as the desktop tool, ported to run client-side with ONNX Runtime Web. View the source on GitHub →