Combination of system and source characteristics for speaker verification under limited data condition

Abstract

There is immense potential for speaker verification system under limited data condition in several real life applications. This paper explains how the combined dissimilar characteristics of voice data improve the performance of speaker verification when training and testing data lengths are reduced (less than 15 sec). To carry out this work, Mel-Frequency Cepstral Coefficients (MFCC), Linear Prediction Cepstral Coefficient (LPCC), Linear Prediction Residual (LPR) and Linear Prediction Residual Phase (LPRP) features are considered. The features are extracted from these extraction techniques are studied individually and pooled them to acquire better verification performance. The experimental evaluation is made by different classifiers of Gaussian mixture model (GMM) and GMM-universal background model (GMM-UBM). The NIST-2003 dataset is used for conducting experiments. The combined dissimilar features provide relatively improved performance compared to all individual features. The GMM-UBM classifier comparatively gives reduced equal error rate (EER) compared to GMM.