Automated assessment of speech disorder severity using global statistics and boosting

This research investigated two novel classifiers designed to identify the level of severity present in a dysarthria speech sample. Severity level refers to the stage of dysarthria, and for this research three levels were identified as mild, moderate, and severe. The first classifier is based on using global statistics of common speech processing features. The global statistics were used to build feature sets that were optimized using a forward sequential algorithm. A maximum likelihood classifier based on a kNN probability function estimate assigns a test sample to one of the three severity levels. This classifier provided accuracy of over 60%. In order to improve performance, a second method called "boosting" was implemented. Boosting is a method that iteratively boosts the importance of some training samples to improve the accuracy of a weak learning system. The specific method of boosting used in the second classifier was adapted from the AdaBoost method. This method only allows for two classes to be analyzed at a time, so a decision tree classifier was implemented to account for all three classes. Results of applying boosting to the original severity classifier show significant improvement of results from both the two-class scenarios and the tree classifier method.