An Online Reversed Bangla Sign Language Learning System

Hearing and speech impaired people around the world express their thoughts through sign language, which is incomprehensible to most people. As a result, interacting with them is quite challenging. Specially in Bangladesh, where people are unable to learn the Bangla sign language (BdSL) due to a lack of resources. We have proposed and implemented an online reversed Bangla sign language learning system using unsupervised rule-based method, which can be utilized to interpret Bangla natural language (speech and/or text) into BdSL. The system takes input of Bangla speech via a Bangla voice recognition module and then convert it into Bangla text. The system can also take Bangla text directly from the text editor module of the system. The system parse Bangla text using a designed set of rules from Bangla semantic analysis. After semantic analysis the system is trained based on unsupervised learning using the parsed Bangla text for each corresponding BdSL. We have used an efficient animation algorithm instead of video clips to display the BdSL for corresponding Bangla speech or text. We have considered only starting and ending frame of each gesture to make the animation. As a result, by using limited number of postures, the system can dynamically make any gestures based on the proposed animation algorithm. The system can animate sentences of various lengths and numerals ranging from 0 to 9 in BdSL. It can generate many phrases with a small quantity of data and has been trained and and tested with over 500 examples with lower computation cost rather than existing systems. The proposed system achieves 99.60% accuracy for speech/voice inputs and 100% accuracy for text inputs where as the average accuracy is 99.8% with the mean computational cost of 120 milliseconds per gesture animation.