Speech recognition for people with dysarthria using convolutional neural network

Dysarthria is a motoric speech impairment caused by neurological impairment. People with dysarthria often find difficulty in moving their muscles, including the ones around mouth and articulators; thus, the speech produced is not too intelligible. Since speakers with dysarthria are often physically incapacitated, Automatic Speech Recognition (ASR) is more preferred to be implemented in an assistive technology than conventional input method such as switch or keyboard. However, commercial ASRs available today have not reached a good performance when being used by speakers with dysarthria. Convolutional Neural Network (CNN) is well-known for its capability at recognizing pattern, including speech. Its implementation in ASR is able to achieve good performance. In this research, CNN is implemented to build a speaker-dependent isolated-word digit speech recognizer for speakers with dysarthria. The recognizer model is built and evaluated with data of 3 speakers with dysarthria and 1 control speaker. Data speech is provided by UA Speech Database. The best performance obtains average accuracy of 90.43 and NRMSE of 0.1366. Overall, not only speech intelligibility affected the performance, but variety of utterances duration might also have impact on how accurate the classification was. © 2019, ICIC International. All rights reserved.