• Version
• Download 1
• File Size 316.09 KB
• File Count 1
• Create Date 3 July 2026
• Last Updated 3 July 2026

Generative AI for Real-Time Translation and Accessibility Solutions

Generative AI for Real-Time Translation and Accessibility Solutions

Muralidharan V ¹, Nishanth S ² , Vaseegaran P ³ , Gill Rinisha k ⁴ , Agesta Jenifer A⁵

¹ PG-MBA & Karpagam College of Engineering – SoMS, Coimbatore, India

² PG-MBA & Karpagam College of Engineering – SoMS, Coimbatore, India

³ PG-MBA & Karpagam College of Engineering – SoMS, Coimbatore, India

                                  ⁴ PG-MBA & Karpagam College of Engineering – SoMS, Coimbatore, India

⁵ UG-CSE & Holy Cross Engineering College, Thoothukudi , India

---------------------------------------------------------------------***---------------------------------------------------------------------

ABSTRACT – India's extraordinary linguistic diversity with 22 scheduled languages, over 19,500 mother tongues recorded in the 2011 Census, and literacy and language proficiency gaps that disproportionately affect rural populations, persons with disabilities, and economically disadvantaged communities creates substantial barriers to equitable access to education, healthcare, government services, and digital platforms. Simultaneously, the World Health Organization estimates that over 1.3 billion people globally live with some form of disability, with visual, hearing, and speech impairments creating accessibility barriers that conventional digital interfaces fail to adequately address. This paper presents a unified Generative AI-based Real-Time Translation and Accessibility (GAITA) framework that integrates multilingual neural machine translation, real-time speech-to-text and text-to-speech conversion, sign language recognition and generation, and generative AI-based content simplification to deliver a comprehensive accessibility solution across linguistic and sensory barriers. The framework employs a fine-tuned multilingual Transformer model (based on the IndicTrans2 architecture) supporting bidirectional translation across 12 major Indian languages and English, integrated with a Whisper-based automatic speech recognition pipeline, a FastSpeech2-based text-to-speech synthesis module with regional voice profiles, and a 3D avatar-based Indian Sign Language (ISL) generation module trained on the ISL CSLRT corpus. A generative AI content simplification layer, built on a fine-tuned instruction-tuned language model, automatically rewrites complex government and educational documents into simplified, accessible language at three configurable reading levels. Evaluation across translation quality, speech recognition accuracy, sign language generation fidelity, and end-to-end latency demonstrates a mean BLEU score of 38.7 across the 12 supported language pairs, word error rate of 8.4% for Indian-accented English and regional language speech recognition, and end-to-end translation-to-speech latency of 1.8 seconds, validating the framework's viability for real-time deployment in classrooms, hospitals, and government service centres.

Keywords — Generative AI, Neural Machine Translation, Speech Recognition, Text-to-Speech, Sign Language Recognition, Accessibility, Indian Languages, Transformer, Multimodal AI, Digital Inclusion