LLM-Enhanced Braille-To-Speech Learning Assistant for Visually Impaired Users Using AraT5

Document Type

Conference Proceeding

Source of Publication

2026 17th Student Research Conference on Applied Computing Src 2026

Publication Date

4-22-2026

Abstract

Braille literacy remains essential for visually impaired individuals to access written information and achieve independence in education and daily communication. However, learning Braille can be challenging due to the lack of immediate feedback when interpreting Braille dot patterns. Traditional assistive technologies often focus only on recognizing Braille characters and converting them into speech without providing contextual explanations that support learning.This paper proposes an intelligent Braille-To-Speech learning assistant that integrates rule-based Braille decoding with the Arabic transformer language model AraT5. The system converts Braille symbols into text, generates contextual explanations using a large language model, and provides auditory feedback through text-To-speech synthesis.Experiments conducted using a Kaggle Braille dataset demonstrate strong performance, achieving 93% Braille decoding accuracy, 0.84 token-level F1 score, ROUGE-1 score of 0.79, ROUGE-L score of 0.76, and BLEU score of 0.71. These results indicate that combining symbolic Braille decoding with large language models significantly improves accessibility and enhances interactive Braille learning for visually impaired users.

ISBN

[9798319510167]

Publisher

IEEE

Disciplines

Computer Sciences

Keywords

Accessibility Systems, AraT5, Assistive Technologies, Braille Recognition, Deep Learning, Large Language Models, Text-To-Speech, Visual Impairment Support

Scopus ID

105043051205

Indexed in Scopus

yes

Open Access

no

Share

COinS