An algorithm using linguistic information and its application to the analysis of speech in the spectral domain

An algorithm is described in this paper which performs spectral analysis of speech and automatic extraction of the formant frequencies from the speech spectra. It performs preliminary classification of the vowels in ´front´,´central´ and ´back´ and processes each of these subclasses according to different rules. The algorithm is programed in FORTRAN-IV IBM 360/370 version. The program was tested by the processing of 614 Bulgarin vowel sounds uttered by 60 speakers/30 male and 30 female/. The performance of the program has shown a total percent of uncorrect decisions about the values of the formant frequencies averaging to 13,25% for the male speakers and 23,60% for the female speakers.