A Lightweight Machine Learning Framework for Real-Time ASL and MSL Recognition

Abstract

For the Deaf and Hard of Hearing (DHH) community, sign language identification is a crucial assistive technology that facilitates communication. Numerous studies have been conducted on American Sign Language, but greater focus is needed on local regional sign languages (MSL). This study emphasizes the importance of a comparative evaluation of ASL and MSL recognition, taking into account distinctions in gesture structure, hand movement and linguistic representation. In this paper a unified real-time identification approach is proposed. We use Random Forest (RF) for classification and MediaPipe for hand landmark extraction. The experimental evaluation shows that the recognition accuracy of both sign languages is high, and MSL has a very high accuracy of 98.87%. These results motivate a more detailed investigation of the reasons behind MSL’s strong performance and the role of linguistic or visual simplicity in the recognition results. The comparative analysis not only demonstrates the feasibility of a unified SLR system in multilingual settings, but also highlights the need to include underrepresented sign languages in accessibility technology solutions.