IMPLEMENTATION OF DNN FOR DYSARTHRIC SPEECH RECOGNITION

Abstract

Dysarthria is the medical condition in which a person loses the control over his motor nerves which are responsible to produce speech. People who are suffering from dysarthria find themselves unable to have a communication with normal people, as people who don’t have any dysarthric person in their acquaintance can not understand the speech of the person. An automatic speech recognition system can help these speakers to convert their unintelligible speech to the intelligible speech signal or text. But the commonly used recognition systems don’t work properly and fail drastically in doing so. In this thesis the issue was addressed and tried to make a recognition system which can recognize the speech of such people at phoneme level. With the use of CNN architecture, a CNN model was made which can interpret the spoken phonemes using the spectrogram. With the accuracy of up to 89% the phonemes are detected and with the help of a linguistic model this architecture can be used to help these patients.