DESIGN AND IMPLEMENTATION OF A STATIC HAND SIGN TRANSLATION SYSTEM

This project presents the design and implementation of a prototype static hand sign translation system that recognizes the 26 static letters of the American Sign Language (ASL) finger-spelling alphabet (A–Z) from live webcam input and converts them into readable, persistent text. The system is implemented in Python and combines CVZone (a high-level wrapper for MediaPipe) for robust, real-time hand detection and 21-landmark extraction with a compact convolutional neural network (CNN) classifier trained on curated static-hand images. This project presents the design and implementation of a prototype static hand sign translation system that recognizes the 26 static letters of the American Sign Language (ASL) finger-spelling alphabet (A–Z) from live webcam input and converts them into readable, persistent text. The system is implemented in Python and combines CVZone (a high-level wrapper for MediaPipe) for robust, real-time hand detection and 21-landmark extraction with a compact convolutional neural network (CNN) classifier trained on curated static-hand images. The operational pipeline first acquires video frames from a standard laptop webcam, uses MediaPipe via CVZone to detect the hand and extract a stabilized bounding box and landmark set, then preprocesses the cropped hand region (resizing, normalization, and optional augmentation) before feeding it to the CNN for per-frame classification. Recognized letters are displayed on-screen and appended to a downloadable text file for easy copying into word processors, enabling a simple assistive communication workflow for finger-spelled input. The implementation intentionally focuses on static poses; dynamic ASL gestures that require temporal motion (notably ‘J’ and ‘Z’) are excluded from the scope. The prototype was evaluated qualitatively using live webcam trials across multiple users, hand sizes, and indoor lighting conditions; results demonstrate consistent, near-real-time recognition of static poses under typical conditions, and reveal common failure modes such as occlusion, extreme low light, and rapid hand motion. Design choices—including landmark-stabilized cropping, light augmentation during training, and a lightweight CNN architecture—prioritize responsiveness on consumer hardware while maintaining competitive classification performance. The project contributes a reproducible, student-level reference implementation and complete Python codebase suitable for further research or deployment. Future work will expand dataset diversity, incorporate temporal models for dynamic gestures, and integrate language-model post-processing to convert 6 letter sequences into validated words and phrases, improving usability for practical assistive applications. By successfully simulating finger-spelling translation in a low-cost, real-time environment, this study contributes to the growing field of AI-driven assistive technologies for the deaf and speech impaired community. It highlights the accessibility of modern hand-tracking APIs and showcases how computer science students can leverage open-source tools to bridge communication barriers. This work therefore represents both a practical demonstration of applied machine learning and a platform for future enhancement in sign-language processing research.