Algorithms for vowel recognition in fluent speech based on formant positions

This paper deals with Czech vowels and two algorithms for vowel detection in fluent speech. Basic algorithm is based on the frequencies of first two formants F1 and F2 determining the individual vowels. The accuracy of basic algorithm is susceptible to the quality of recorded speech signal because voice distortion and surrounding noise can cause a high rate of false vowel detection. Second algorithm is an improved version on the first one extended by retroactive checking of vowel duration. Both algorithms were applied on created speaker database containing 5 female and 13 male native speakers. Better results were obtained using the improved algorithm which reduced false detection of vowel segments over all speakers and all vowels by 36.7% at average.