Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Phoneme Recognition in Fixed Context Using Regularized Discriminant Analysis

A. Rudzionis, V. Rudzionis

Kaunas University of Technology, Kaunas, Lithuania; and Vilnius University, Vilnius, Lithuania

Speaker independent discrimination of four confusable consonants in the strictly fixed context of six vowels is considered. The consonants are depicted by features of consonantís stationary part and changing rate of features (delta features) in transition from consonant to the following vowel. The mel frequency cepstrum (MFCC), linear prediction cepstrum (LPCC), recursive filter (F12) features and set of discriminants were evaluated seeking for better phoneme discrimination. It is postulated that Gaussian mixture capabilities are similar to k-means (kMN) capabilities and several discriminants including regularized discriminant analysis (RDA) were analyzed too. The experiments showed that the discrimination error averaged per environments of six vowels decreases from 23.3% using kMN to 7.0% using RDA for the best F12 features. Consonant discrimination error rate decreases from 21.6% to 3.6% in the open vowel context and from 27.9% to 11.4% in closed vowel context.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Rudzionis, A. / Rudzionis, V. (1999): "Phoneme recognition in fixed context using regularized discriminant analysis", In EUROSPEECH'99, 2745-2748.