Automatic Early Detection of Dysarthria using Deep Neural Network

Dysarthria is dis-coordination among articulatory muscles responsible for articulation in the production of speech. A person with dysarthric disorder cannot control their tongue, hence, are prone to slurred speech or swallowing words which lead to improper pronunciation and less intelligibility of speech. This work focuses on the automatic early detection of Dysarthria. To elucidate the main idea, it must be understood that the whole detection procedure revolves around the speech intelligibility of a person. It becomes an arduous job to distinguish between a healthy person's speech and a person who is slowly evolving with the disorder. The prime motive of the work is to provide a proper classification of the disorder in its earliest stages. The common words and phrases recorded in the Universal Access database are used for this work. Mel Frequency Cepstral Coefficients (MFCC) are extracted from both controlled and high intelligible speech data, and fed to a Deep Neural Network (DNN) classifier. MFCC fed to the DNN classifier has provided the best classification accuracy for the above experimental setup.