Vowel Identification Using Piecewise Separation Technique
Debabrata Majumder · 1978
A simple method for computer recognition of Telugu speech sounds irrespective of speakers is described. A vocabu bulary consisting of 871 Telugu words containing the ten vowels (f'O/./a :/,/i/,/i :/,/u/,/u :/,/e/,/e :/./0/ and /0 :/) in constant vowel nucleus-consonant (CNC) combination and uttered by three informants was selected as the testing material. Formant frequencies. Flo F2 and F3 of the vowel sounds were extracted by spectrographic analysis. Piecewise using (F 2 -F1) or (F3/F2) as parimary recognition parameters and (F3/F,) as a final recognition criterion were used for classification. From frequency distributions of these parameters for both shorter and longer vowel categories suitable boundaries have been selected. III the classification model. shorter and longer subgroup of a vowel have been pooled together in the same class and the overall recognition score is about 84 %. MANY attempts have been made to build a speech recogn.ition system l - 5 such that vowel speech sounds could be recognized in almost all cases either from the conditional samph::d data of speech waveform 2 • 6 or from the vowel formant frequencies with a suitable discriminant fUllction 7 ,8. In. the second case, the acoustic parameters of vowels such as voice funda mental frequency (Fo) and formant frequencies (FJ, F2 , F3, etc.) depend greatly on sex and age. It has been found that o and F3 do not bear as much of the information about phoneme identity as do 1 and F2; they are also speaker dependent and possess close correlatio:1. with age and sex. Every measure ment on a vowel in (Fo-F3) plane corresponds to almost identical region irrespective of the kind of vowel9 . For this reason, vowel identification for a number of informants does Show an improved result when either F3 or Fo or both F3 and Fo in addition to F( and 2 are taken into cO:l.sideration. In the present paper we have made an attempt to wards the recogllition of Telugu vowels irrespective of speakers by considering only the third formant F3 in addition to F1 and 2 • The classification tech nique is based on piecewise separation by a decision boundary determined from the distributiOll functions of the characterizing parameters. These parameters were found from the interrelatioas among F F2 and F3 · For threshold boundaries, (F 2 -F)) or (F3/F2) were taken as the recognizing criteria in primary separation and (F3/F1) is the Jast criterion for final recognition. Their distribution functions after pre processing were also studied. Finally, a suitable piecewise linear classification scheme with the mini mum number of errors is described. The results of computer identification of vowels are presented in tabular form.