SYSTEM ARCHITECTURE AND ACCESSIBLE INTERFACE DESIGN FOR REAL-TIME UZBEK SIGN LANGUAGE TRANSLATION

Abstract

Real-time sign-language translation systems are usually evaluated on recognition accuracy alone, but accuracy is only useful if the surrounding system actually reaches its intended users. This paper describes the system architecture and accessible interface design behind our Uzbek Sign Language (UzSL) translation prototype, developed as part of a broader sign-language translation project [1]. We separate video capture from model inference across a client-server architecture. Frames stream from an in-browser client to a FastAPI backend over WebSockets [4], and the recognition model is served through ONNX Runtime or PyTorch and deployed in a Docker container [5, 6] for portability across environments. The web interface, built with Next.js, Tailwind CSS, and Radix UI [7], was evaluated against the Web Content Accessibility Guidelines (WCAG) 2.1 [2], including screen-reader compatibility, keyboard-only navigation, and colour-contrast requirements. We report a series of refinements made in response to field observations during testing, including nighttime-recording support for low-light conditions, contextual thresholds to reduce confusion between visually similar signs, a caching layer that cut speech-synthesis latency by roughly 40% [3], and interface adjustments such as larger tap targets for younger and older users. We present these architectural and accessibility decisions as a practical reference for deploying real-time recognition systems for low-resource sign languages, independent of the specific recognition model used.