Deep learning based Braille character to Bangla voice conversion system

Abstract

Braille is an essential mode of communication for visually impaired individuals, enabling them to read and understand contexts through a tactile system of raised dots. However, certain regions worldwide, especially Bangladesh, have limited access to digital solutions for Braille conversion, which poses a challenge for those individuals. This thesis presents a method for converting Braille Characters to Bangla Voice through a deep-learning based system and thus enhancing accessibility for visually impaired individuals as well as general individuals who cannot understand Braille. In this research we have addressed the scarcity of high volume labeled data consisting of more than 10000 high-resolution labeled data with approximately 1 million labeled braille instances by strictly maintaining Library of Congress spatial standards. This research simulataniously performs two different architecture one being a multi-stage segmentation-classification and other one being YOLO based unified architecture. The unified architecture vastly outperforms the traditional method by achieving Mean Average Precision (mAP@50) of 98.5 percent in character extraction. The recognized text is then converted into natural-sounding Bangla speech using a TTS engine. This system is evaluated through accuracy, processing speed, and friendly user experience, demonstrating its potential to bridge the communication gap for the visually impaired in Bangla-speaking communities.