Significance of Entropy Based Features For Dysarthric Severity Level Classification

Dysarthria is a motor speech disorder arising from impairment of muscles that makes difficult to form or pronounce words while speaking. In this paper, we introduce an approach of multiband entropy based features extracted from dysarthric speech signals for dysarthric severity level classification. Generally, entropy is measured as the number of bits of information contained in each message signal. The information content of these signal measures how much randomness or uncertainity contains in a signal. Extending this, we use a frame-wise processing technique to divide the speech signal into short frames which allows for detailed analysis of different characteristics of speech signal. Furthermore we divide frames into equal sub bands and compute the entropy and zero mean entropy in each sub band using Gabor filterbank. It is expected that the mean entropy of very low dysarthric severity is lower compared to high dysarthric severity level indicating that randomness is increasing as the severity level of dysarthria is increasing. Experimental analysis were conducted on extensively used dataset namely UA Speech. Results were carried out by Convolutional Neural Network (CNN), along with 5-cross validation. The results are compared against standard MFCC, LFCC, and glottal source based LFRCC. It was observed that the addition of entropy information boosted the performance of MFCC by 4.38%, LFCC by 2.54%, and LFRCC by 1.88% compared to their traditional techniques indicating the crucial information captured by the entropy for dysarthria severity classification.