Age and Gender Recognition of a Speaker from Short-duration Phone Conversations

Abstract

In this study, a system is suggested that is classifying phone conversations having an average length of one second into three categories according to the age and/or gender properties of speakers. In the study where SVM approach based on GMM supervectors is used, speech signals are represented by an 39-element vector consisting of MFCC coefficients. In the tests where aGender database is used, the effect of GMM component number on the success is also investigated and the most suitable component number is determined. At the end of these tests, for the three-class gender category 87.28%, four-class age category 50.12% and seven-class age&gender category 48.05% success rates are achieved.