Beyond Letter Accuracy: A Real-Time Bangla Sign Language Fingerspelling System with Gesture-Driven Word Construction and Lexicon-Based Correction

This work presents a real-time Bangla Sign Language fingerspelling system that goes beyond isolated letter recognition by integrating gesture-driven word construction and lexicon-based correction. The system combines hand landmark processing, CNN-based character recognition, temporal filtering, and user-controlled gestures to form complete words from sequential signs. A lexical correction stage is incorporated to reduce errors in word formation and improve practical usability. The approach was evaluated using the BDSL49 dataset and achieved approximately 97% validation accuracy for character recognition. The study focuses on building a more usable end-to-end fingerspelling pipeline rather than treating sign recognition as an isolated classification task.