Dysarthria Speech Disorder Classification Using Traditional and Deep Learning Models

Abstract

Dysarthria is a motor speech disorder that results in speech difficulties due to the weakness of associated muscles. This unclear speech makes it difficult for dysarthric patients to present himself understood. This neurological limitation is usually occurs due to damages to the brain or central nervous system. Speech therapy can be effectively employed to enhance the range and consistency of voice production and improve intelligibility and communicative effectiveness. Assessing the degree of severity of dysarthria provides vital information on the patient's progress which inturn assists pathologists in arriving at a treatment plan that includes developing automated voice recognition system suitable for dysarthria patients. This work performs an exhaustive study on dysarthria severity level classification using deep neural network (DNN) and convolution neural network (CNN) architectures. Mel Frequency Cepstral Coefficients (MFCCs) and their derivatives constitute feature vectors for classification. Using the UA-Speech database, the performance metrics of DNN/CNN based learning models have been compared to baseline classifiers like support vector machine (SVM) and Random Forest (RF). The highest classification accuracy of 97.6\% is reported for DNN under UA speech database. A detailed examination of the performance from the models discussed above reveal that appropriate choice of deep learning architecture ensures better results than traditional classifiers like SVM and Random Forest.