Realtime conversation system for people with hearing and speech impairments

Abstract: Communication barriers significantly hinder interaction between the deaf community and the wider world. This paper investigates an automatic system for Indian Sign Language (ISL) detection using MobileNetV2, a transfer learning architecture known for its efficiency. We leverage transfer learning from pre-trained MobileNetV2 weights to extract features from ISL images. To improve model performance for ISL detection, we incorporate linear bottleneck layers and squeeze-and-excitation blocks within the network. Additionally, separable convolutions are used to maintain accuracy while reducing computational complexity. This optimized MobileNetV2 architecture is then fine-tuned on a prepared ISL dataset for robust sign recognition. While limitations exist, this research paves the way for advancements in communication accessibility for the deaf community.

Keywords: Indian Sign Language (ISL),Sign Language Detection ,Deep Learning,MobileNetV2,Transfer Learning ,Linear Bottleneck Layers ,Squeeze-and-Excitation Block ,Communication Accessibility ,Deaf Community.