Selection of features and classification rules for Slovene phone recognition

Results are presented of testing different feature sets and classification rules for Slovene phone recognition purposes. Different parametric descriptions of speech such as Fourier, cepstral, and linear prediction analysis of 12.8-ms long overlapping signal frames were used. It was found that linear filter spacing, gives significantly better recognition results for some phone categories than logarithmic spacing. Recognition results using LPC-cepstrum parametric representation were not significantly different from those obtained by MEL-cepstrum when using the Bayes classifier assuming a normal distribution of the feature vector.