Consonant Class Discrimination in Dysarthric Speech Based on Support Vector Machine Using Class- Dependent Acoustic Parameters

In this paper, we propose a consonant class discrimination (CCD) method in dysarthric speech, where a support vector machine (SVM) is employed by using class-dependent acoustic parameters. To this end, each consonant is categorized into one of five classes according to the manner of articulation such as stop, affricate, fricative, nasal and glide. In the proposed CCD method using SVM, acoustic parameters distinctive to consonant classes are extracted and they include the spectral energy ratio between the low and high frequency bands, zero crossing rate, spectral flatness, spectral peak frequency, and so on. The effectiveness of the proposed CCD method is demonstrated by using a database composed of mild and severe dysarthric speeches. As a result, it is shown that the proposed CCD method employing class-dependent acoustic parameters relatively reduces average discrimination error rate by 7.67%, compared to that employing mel-frequency cepstral coefficients.