Comparing ANN, SVM, and HMM based Machine Learning Methods for American Sign Language Recognition using Wearable Motion Sensors

Abstract

Millions of people with speech and hearing impairments, worldwide, communicate through sign languages every day. In the same way that voice recognition provides a simple communication platform for most users, gesture recognition is a natural means of correspondence for the hearing-impaired. In this paper, we explore the problem of translating/converting sign language to speech, and propose an improved solution using different machine learning techniques. We seek to build a system that can be employed in the daily lives of people with hearing impairments, in order to enhance communication and collaboration between the hearing-impaired community and those untrained in American Sign Language (ASL). The system architecture is based on using wearable motion sensors and machine learning techniques. In this study, we propose a solution using Artificial Neural Networks (ANN) and Support Vector Machines (SVM), and compare their accuracy with the Hidden Markov Model (HMM) results from our previous work to recognize ASL words. Experimental results show that using ANN gives an overall higher accuracy in recognizing ASL words, compared to other machine learning techniques.