AI Workflow · Mar 2026
SignStream - Real-Time Browser-Based Sign Language Recognition
Client-side real-time sign language recognition (SRL) with MediaPipe landmarks, SignGRU, and ONNX - plus personal fine-tuning.
About
Built a browser-based ASL recognition app with live caption buffering and optional text-to-speech (Web Speech API) for accessibility. Trained a bidirectional GRU (BiGRU) classifier on 16-frame MediaPipe hand-landmark sequences using WLASL (top-100) plus ASL alphabet data (129-class vocabulary); achieved 97.2% ± 0.25% top-1 accuracy (5-fold stratified CV). Deployed a fully client-side inference pipeline using MediaPipe and ONNX Runtime Web (no backend inference), enabling real-time predictions and caption display in the browser. 📄 Research paper presented at NCMASE-26 (Mar 2026): "SignStream: A Client-Side Real-Time Browser-Based Sign Language Recognition System"