Analysis and development of clinically recorded dysarthric speech corpus for patients affected with various stroke conditions

Abstract

The manuscript presents the work related to the development of a dysarthric speech corpus for various types of stroke conditions. The corpus consists of speech recorded from 50 stroke patients and 50 healthy controls in clinical environments. Severity of stroke for each patient has been assessed by the clinician based on the National Institute of Health Stroke Scale. The text read by patients and healthy controls comprises (a) five sustained vowels, (b) three words consisting of the plosive consonant and vowels, and (c) 10 phonetically rich sentences in Telugu language. A discriminative analysis is carried out using conventional Mel Frequency Cepstral Coefficients and Convolutional Neural Networks to quantify the perceptual variations in dysarthric speech of stroke patients and healthy controls. Vowels and word utterances of the speech corpus exhibited better class discrimination characteristics compared to sentences for text dependent and speaker independent scenarios.