VOICE BASED GENDER IDENTIFICATION SYSTEM

Abstract

The purpose of this research was to evaluate and compare the gender identification performance of two approaches, Gaussian Mixture Model - Universal Background Model (GMM-UBM) and Gaussian Mixture Model - Support Vector Machine (GMM-SVM). The Mel-frequency cepstral coefficients (MFCC) was used for this comparison. Experimental results were conducted on a British English speech corpus named the 'Accents of the British Isles' (ABI-1), it is shown that the gender identification performance using the GMM-SVM system is significantly better than the GMM-UBM system. Our results claimed that the accuracy of identifying the gender of new speakers using the GMM-SVM system can reach up to 99.44% and up to 96.99% using the GMM-UBM system. Also, another study based on utterances of short length proved that the GMM-SVM system still capable of identifying unseen data with high accuracy up to (99%) compared to the GMM-UBM system (97%). Experiments on the effect of utterance length, number of MFCCs and number of GMM components for both systems are discussed in this dissertation.