Correlogram Based Feature Selectıon For Speaker Identıfıcatıon Usıng Vowels

Abstract

A correlogrcnn approach to the seîection of text-dependent feaîures mvcweî souncis is invesiigaîed for speaker identificaîion. in the approach, vowel soundsas the idenîity carrying parts in spoken utterances are represented m the form of acorrelogram, m whıch the speaker dependent spectral and temporaî information iscodsd. Psycho-physiologıcally motivated speclro-temporal correlation with a searchalgoriihm is introduced to identify the regions vhere the relevanl features areembedded thal are suited to discrimination. We identify Ihefeature regionsfor a set ofındividuaî vowel sounds, and present resulîs on their effectiveness in identifyingspeakers. Partıculccr to îhe approach is thaî U makes no explicit use of any individîialspeech features.

Özet

A correlogrcnn approach to the seîection of text-dependent feaîures m