Research

This project builds and evaluates a real time American Sign Language recognition desktop app that runs fully on a user's computer. The system uses an Electron based front end and a Python backend connected through WebSockets. A single deep learning model, trained on the MS ASL dataset of 222 signers, recognizes word level ASL from live webcam input. The model uses a CNN backbone with a temporal layer for short sequences of frames, so it can handle motion and variation in signing speed. The project focuses on making this one architecture accurate, robust, and fast enough for interactive use, then measuring its performance across different users, lighting conditions, and frame rates.