Speaker distinction using vowel polygons: Experimental study

Abstract

This paper presents experimental study oriented on finding the most suitable polygon generated by Czech vowels for speaker recognition. For created speaker database containing 37 Czech native speakers, speaker variances have been observed by the length of vector created by two different centers of gravity for 16 different vowel polygons in 10 formant planes. The suitability is regarded by the dispersion coefficient of generated distance vectors and their minimal value. The most suitable vowel polygons have been set as EIOU25 tetragon and IOU34 vowel triangle reached the best results in both criteria. The dispersion of individual centers of gravity is much more suitable for speaker recognition generated by higher formants. Presented observations can be further used in the case of actual emotional state recognition or alcohol/drug detection.