Speaker Verification in Different Database Scenarios

Abstract

Abstract. This document shows the results of our Speaker Verification System under two scenarios: the Face and Speaker Verification Evaluation organized by MOBIO (MObile BIOmetric consortium) and the results for the Speaker Recognition Evaluation 2010 organized by NIST. The core of our system is based on a Gaussian Mixture Model (GMM) and maximum likelihood (ML) framework. First, it extracts the important speech features by computing the Mel Frequency Cepstral Coefficients (MFCC). Then, the MFCCs train genderdependent GMMs that are later adapted to obtain target models. To obtain reliable performance statistics those target-models evaluate a set of trials and final scores are calculated. Finally, those scores are tagged as target or impostor. We tried several system configurations and found that each database requires a specific tuning to improve the performance. For the MOBIO database we obtained an average equal error rate (EER) of 16.43 %. For the NIST 2010 database we accomplished an average EER of 16.61%. NIST2010 database considers various conditions. From those conditions, the interview training and testing conditions showed the best EER of 10.94 %, followed by the phone call training phone call testing conditions of 13.35%.