A modern approach to dysarthria classification

This work deals with the assessment of neurological diseases known as dysarthrias, using a novel approach based on objective and perceptual features extracted from pathological speech signals. A methodology for the classification of dysarthria is developed in which digital signal processing algorithms are used to appraise the severity of those features less reliably judged by the clinicians, while the others are taken directly from perceptual judgments or medical records. The assessment process evaluates the performance of two different classifiers and compares them with the traditional assessment system. The first approach is based on the lineal discriminant analysis and the second is a non-lineal technique based on self-organizing maps. The non-lineal classifier provided the highest percent of correct classification and the most accurate information on the relevance of the features in the classifier decision. It also provided a bi-dimensional representation of de data that allows a better understanding of the correspondence between the speech deviations and the location of the damage in the peripheral or central nervous system.