SignBridge v2: An Improved Transformer Architecture for Real-Time Bidirectional ASL Communication

Abstract

SignBridge v2 is a real-time, bidirectional American Sign Language (ASL) communication platform designed to bridge communication between ASL users and non-signing individuals. The system performs sign-to-text and speech translation using MediaPipe hand landmarks and a custom Transformer-based architecture, SignTransformer, trained to classify 250 dynamic ASL signs. The v2 model achieves 80.24% Top-1 accuracy and 92.46% Top-5 accuracy on the held-out test set. The platform also converts typed or spoken input into ASL skeleton animations using a compact keypoint database, reducing storage requirements by 99.2% compared with video-based approaches. The system is browser-based, hardware-independent, and designed around a stateless architecture with no persistent storage of user data.