Bridging Communication Barriers: A Review on SLR by means of Machine Learning and AI

The sign language is a vital communication medium for Deaf and Hard-of-Hearing people. But it is difficult for normal people to understand it as they are not familiar with its actions or movements. It is necessary to translate the sign language gestures into normal speech so that normal people can communicate with Deaf and hard of hearing peoples or signers. It is made possible by using Sign Language Recognition (SLR) systems that are driven by machine learning (ML) and artificial intelligence (AI). This review work considers vision-based, sensor-based, and hybrid practices of SLR and checked the progress of recent work in this field. In order to enhance gesture identification, feature extraction, and sequence modelling, the functions of deep learning (DL) designs including Transformers, Recurrent Neural Networks (RNNs), and Convolutional Neural Networks (CNNs) are investigated. These includes discussion of issues like occlusion, background noise, signing style heterogeneity, and the lack of annotated datasets. The study also assesses benchmark models, performance indicators, and existing datasets while outlining present shortcomings and potential avenues for further research. This paper assesses the SLR systems (Vision, sensor and hybrid) on the basis of some performance metrics like: Input modalities, Accuracy, Robustness to environment, Real-time performance and Intrusiveness. It was observed that vision-based systems work well in case of non-intrusive settings as this system do not need wearable devices, while the sensor-based systems are highly intrusive due to wear hardware. In case of uneven environmental conditions, the accuracy of vision-based system drops. On the other hand, accuracy of sensor based and hybrid model remain high in all conditions. Vision based systems have reasonable real time performance, sensor based and hybrid systems are faster for real time applications.