Mexican Sign Language Corpus: Towards an Automatic Translator

Abstract

The development of the Sign Language Corpus has been motivated by its great utility and application to various purposes and research areas. However, some countries do not have their own Sign Language Corpus. Developing a corpus thereby benefits the community of people with speech disabilities in diverse areas such as education. Thus, the motivation to develop this work is to present an advance toward constructing an RGB-D corpus of Mexican Sign Language captured by a Kinect sensor. A total of 90,000 samples of 570 words and 30 phrases interpreted by 150 people who commonly use Mexican Sign Language were collected. Of the participants, 86 were women and 64 were men, aged between 12 and 60 years old. The Mexican Sign Language Corpus was recorded by signers from three different regions of the south of Mexico. The constructed corpus contains depth, color, point clouds, and human skeleton positions. Six hundred of the most used words were selected from 17 semantic fields, considering the variability in the movement of both hands. After training a neural network, the performance developed by the recognition system was 98.62%.