Analysis of Features for Dysarthria Severity Classification from Speech

Dysarthria is a disorder that affects the ability of an individual to speak clearly. Diagnosis of the severity of dysarthria can aid in providing appropriate therapy to the individual. Therefore, the current work focuses on identifying features that can be used to estimate the severity of dysarthria. In this regard, two feature sets are considered, namely DisVoice and OpenSMILE, and the genetic algorithm is used to identify the most suitable list of features from both feature sets using 3 speech corpora, namely SSN-TDSC, Torgo, and UA Speech corpora. In order to test the effectiveness of these features, classifiers are trained on each feature set as a whole and on the features identified by the genetic algorithm, and their performance compared. It is observed that articulatory and prosodic features are most important in the diagnosis of severity of dysarthria. A maximum accuracy of 85% is obtained with the shortlisted DisVoice features and 97% with the OpenSMILE features.