Abstract
Although Support Vector Machines (SVMs) have been proved to be very powerful classifiers, there performance still depending on the features representations which they used. This paper describes an application of SVM to multiclass phoneme recognition using 7 sub-phoneme units and different features representations as MFCC, PLP and RASTA-PLP. The phoneme recognition system is tested and experimentally evaluated using speech signals of 49 speakers from TIMIT Corpus in order to find the adequate feature coefficient for our phonemes databases. The experimental results show that, MFCC and PLP are significantly superior and get better recognition rates than RASTA-PLP (52% vs. 29%).