Vowel Recognition Using a Multiple Discriminant Function

An application of statistical detection theory to automatic speech recognition is presented. The multiple discriminant function approach is applied to the problem of vowel recognition. Twenty-one speakers, 10 females and 11 males, produced 215 single vowel words, exemplifying ten different vowels. The linear discriminant function correctly labeled 87% of the utterances; the quadratic discriminant function correctly labeled 94%. It is pointed out that these detections were achieved without use of duration information and without explicit use of formant information.