Speech Pathology Classification

Neurological diseases, such as Alzheimer's, Parkinson's, or organic disorders, affect speech production. Sometimes, patients have difficulty recognizing a change in their voice. It is necessary to classify voice disorder severity automatically to diagnose, monitor, or prevent voice disorders. The paper is devoted to dysarthric speech classification as a case study using the UA-Speech benchmark dataset. Four recording classes are rated with respect to intelligibility: very low, low, medium, and high. Two classification systems are compared. The first system is a distance-based classifier that exploits the identity vectors (i-vectors) extracted from speech recordings. The second system is an end-to-end deep learning system resorting to Spatial Convolution Neural Networks applied to spectrograms.