AVERAGED TEMPLATES CALCULATION AND PHONEME CLASSIFICATION

Abstract

Modeling of acoustic processes has not been considered sufficiently in experimental research of phoneme classification. This paper presents phoneme template calculation method which is motivated by modeling of transitional and stationary phonetic processes and enables to represent each of the distinctive process by a single feature vector. Experimental study confirmed advantages of the proposed method. The best achieved phoneme recognition accuracy was 65.2% in our study. This result is by 26.4% better than the recognition accuracy achieved with phoneme representation composed from all frames composing phoneme. It was observed that the best vowel, semivowel and fricative consonant representation were got from 40-50 ms while best representations of plosive consonants were got from 30 ms.