Federated Deep Learning for Bidirectional English and American Sign Language Translation Across Distributed Deaf Communities

Communication barriers between Deaf and hearing individuals continue to limit accessibility across educational, healthcare, workplace, and community settings. Existing Artificial Intelligence (AI)-based translation systems often focus on isolated sign recognition and rely on centralized datasets with limited signer diversity. This paper presents a Federated Deep Learning (FDL) framework for bidirectional translation between spoken English and American Sign Language (ASL). The proposed system integrates computer vision, transformerbased neural machine translation, and federated learning to enable privacy-preserving collaborative model training across multiple institutions while accommodating regional and signerspecific ASL variations. The framework was evaluated using data collected from five collaborating institutions involving 412 participants and more than 68,000 annotated ASL video samples. Experimental results achieved translation accuracies of 91.3% for ASL-to-English and 88.7% for English-to-ASL translation. Compared with a centralized transformer baseline, the proposed federated model improved BLEU scores by 7.8% while enhancing generalization across signing communities. User evaluations involving Deaf educators and community members produced a System Usability Scale score of 86.4, indicating excellent usability. These findings demonstrate the potential of federated learning to support scalable, privacy-preserving ASL translation systems and improve accessibility technologies for Deaf communities.