Enhancing Communication Accessibility: A Deep Learning Approach for Assistive Hand Gesture Recognition in Speech Disability Communities

Individuals with speech disabilities face significant barriers in daily communication, leading to social isolation. While augmentative and alternative communication (AAC) devices exist, they often lack intuitiveness and real-time performance. This study investigates the efficacy of deep learning-based object detection models to create a non-intrusive, vision-based hand gesture translator. A custom dataset of 275 images representing five essential gestures (“Hello,†“Thank You,†“Yes,†“No,†and “I Love You†) was constructed and used to train three state-of-the-art architectures: SSD MobileNet V2 FPNLite 320x320, SSD ResNet50 V1 FPN 640x640 (640x640), and EfficientDet D0 512x512. Performance was evaluated based on precision, recall, and loss metrics. The SSD MobileNet V2 model demonstrated superior performance with a precision of 0.8869 and a recall of 0.8867, offering an optimal balance between accuracy and computational efficiency. In subsequent real-time prediction tests on 100 samples, the system achieved an overall accuracy of 95.5%. The results underscore the potential of lightweight deep learning models in developing affordable and efficient assistive technologies, providing a robust foundation for enhancing communication accessibility and social inclusion.