Comparative study of hybrid models for robust speaker recognition task

Abstract

This paper deals with text-independent speaker verification system based on spoken Arabic digits in real environment. In this work, we adopted Mel-Frequency Cepstral Coefficients (MFCC) as the speaker speech feature parameters, Gaussian Mixture Model (GMM) are used for modeling the extracted speech feature and training the support vector machines (SVMs). The experiments were conducted on the ARADIGIT database at different Signal-to-Noise Ratio (SNR) levels and under two noisy conditions issued from NOISEX-92 database. The obtained results show that the GMM-SVM model outperforms the GMM-UBM, especially in noisy environments.

Keywords: Speaker verification, MFCC, GMM-UBM, GMM-SVM, Noisy environment