On Machine Recognition of Telugu Speech Sounds

Abstract

The paper presents some results on computer recognition of Telugu speech sounds selected from 360 multisyllabic words spoken by 5 male informants. The data are procured from spectrographic analysis of these samples. The basic features considered for vowel recognition are the first three formant frequencies and for plosive consonant recognition are the amount and duration of transition of formant frequencies. For vowel recognition, the score obtained is 83%. The back and front vowels show better recognition scores (92–94%). The average scores obtained for unaspirated plosive recognition are 76% and 57% for CV and VC combinations respectively. For CV combination, duration of transition is found to be an essential feature for classification. Velars and labials, the phonemes of articulatory end positions have a better recognition scores for both CV and VC combinations. A prior knowledge of the target vowels is found to be necessary for attaining a reasonable recognition score. The recognition scores of voiced plosives are better than those of unvoiced one. The maximum likelihood method is used for classification of the vowels and consonants. Two simplifying assumptions of maximum likelihood method can reduce computation times by as much as a factor of five while producing no significant change in recognition accuracy.