Recent research works rely on machine learning models in many speech assistance systems. Machine learning based speech assistance models mainly contributes in transforming dysarthric speech to normal speech will be of great help to persons suffering with this aid. For an accurate speech transformation, best set of features need to be extracted from dysarthric speech and machine learning based classifiers need to be trained with those features for translating to normal speech. Present work does a comparative analysis of feature extraction methods for Kannada bi-syllable dysarthric speech. A clustering-based analysis is conducted on feature extraction methods, each separately and in combination is done.

Through analysis, best feature set combination suitable for accurate recognition of Kannada dysarthric bi-syllable is identified. While earlier works focused feature analysis only based on classification accuracy, But this present work does cluster analysis to calculate the inter distance between the bi-syllables and identify the region where marginal errors can occur in recognition. MFCC, LPC, PLP, LPCC, PE-SFCC, Prosodic features are the feature extraction methods were analyzed and the combination of the feature extraction methods is compared. The clustering based analysis results that the combination of PE-SFCC + LPC + PLP is found to perform better than other feature extraction methods.