Significance of Energy Features for Severity Classification of Dysarthria

Abstract

Dysarthria is a neuro-motor speech disorder that affects the intelligibility of speech, which is often imperceptible depending on its severity-level. Patients’ advancement in the dysarthric severity-level are diagnosed using the classification system, which also aids in automatic dysarthric speech recognition (an important assistive speech technology). This study investigates presence of the linear vs. non-linear components in the dysarthic speech for severity-level classification using the Squared Energy Cepstral Coefficients (SECC) and Teager Energy Cepstral Coefficients (TECC), which captures the linear and non-linear production features of the speech signal, respectively. The comparison of the TECC and SECC is presented w.r.t the baseline STFT and MFCC features using three deep learning architectures, namely, Convolutional Neural Network (CNN), Light-CNN (LCNN), and Residual Neural Network (ResNet) as pattern classifiers. SECC gave improved classification accuracy by $6.28\%$ ($7.89\%$/$3.60\%$) than baseline STFT system, $1.7\%$ ($4.23\%$/$0.99\%$) than MFCC and $0.1.41\%$ ($0.56\%$/$0.28\%$) than TECC on CNN (LCNN/ResNet) classifier systems, respectively. Finally, the analysis of feature discrimination power is presented using Linear Discriminant Analysis (LDA), Jaccard index, Matthew’s Correlation Coefficient (MCC), F1-score, and Hamming loss followed by analysis of latency period in order to investigate practical significance of proposed approach.

This work was done when Dr. Hardik B. Sailor was at Samsung Research Institute, Bangalore, India in collaboration with DA-IICT, Gandhinagar.