Computer-aided recognition of polish vowels in continuous speech

Downloads

Authors

  • Wiktor JASSEM Institute of Fundamental Technological Research Polish Academy of Sciences, Poland
  • Danuta GEMBIAK Institute of Fundamental Technological Research Polish Academy of Sciences, Poland
  • Andrzej DYCZKOWSKI Department of Computer Techniques, Mickiewicz University, Poland

Abstract

Vowels pronounced by male voices in typical Polish sentences were the object of the recognition. Vowel formants as variable time functions were measured from the spectrograms. In the first experiment, the patterns for each phoneme in the form of two-element mean vectors and the appropriate co-variance matrices were averaged over various combinations of 10 voices. In the second experiment the patterns were averaged separately for each of 10 persons. Quadratic and linear discriminant functions were used for the recognition. In general, the recognition scores in experiment I reached 75 % and in experiment II — 90 %. It is assumed that in the automatic recognition of Polish vowels in a computer-aided man-machine system and using two features, high scores may be obtained. They are improved by an adaptation of the system to the operator's voice.

References

[1] T. CALIŃSKI, A. DYCZROWSKI, Z. KACZMAREK, Identification of observation using a dividing hyperplane [in Polish], Rocznik Akademii Rolniczej w Poznaniu, Algorytmy biometryczne i statystyczne, No 4 (1975).

[2] G. I. GJEMEL, Recognition of speech signals [in Russian], izd. Nauka, Moskwa 1971.

[3] M. DIERKACZ, R. GUMIECKIJ, Ł. MISZIN, M. OWIERCZENKO, M. CZABAN, The perception of speech in recognition models [in Russian], izd. Iwowskowo Uniwersiteta, Lwow 1971.

[4] H. DUDLEY, S. BALASHEE, Automatic recognition of phonetic patterns in speech, JASA, 30, 721-732 (1958).