Speech Recognition of Hearing Impaired Children

Abstract

In this world of digital technologies, such as artificial intelligence and voice-driven ap plications, Automatic Speech Recognition (ASR) is becoming essential to our daily life activities specially for disabled people. Now a days, people who are hearing impaired use lipreading visual cues for speech therapy. For our research work, data was acquired from speech therapy sessions of hearing- impaired children. In these sessions, hearing impaired children learnt how to articulate English words through speech therapy. Lan guage therapy of single words as well as pronunciation of short English sentences were practiced by lip reading visual cues. The improvement in speech therapy sessions can be made possible by checking the accuracy of articulated words. The goal of our research is to build speech recognition engine for abrupt and disrupted speech of hearing impaired. We have checked how accurately those words were articulated. Depending on the ac curacy achieved by recognition results, therapists can make hearing impaired children practice those words again which are not correctly pronounced in prior sessions. Hear ing impaired individuals have their own way of pronouncing the words of English due to the abrupt and disrupted speech of hearing impaired. Hence their language model (LM) and dictionary is also unique. We built a language model (LM) for disrupted and cumbersome speech of hearing impaired. We have checked the impact of different model sizes of acoustic model on accuracy of isolated recognized words. We used Kaldi speech recognition toolkit for building our system. In this model, we have checked the perfor mance measure that is, Word Error Rate (WER) based on isolated word recognition. The accuracy of 61% was achieved by the proposed recognizer. We have validated our system performance in comparison with the Google Cloud speech-to-text API built for normal hearing with speech data of hearing impaired. The difference of 45% accuracy in existing work was observed. The recognition results can be further improved with increase in data size. Graphical user interface can be built in future to facilitate hearing impaired children and instructor