Tests of robustness of GMM speaker verification in VoIP telephony

Abstract

The paper presents the scores of the GMM (Gaussian Mixture Models) based speaker verification system for speech signal transmitted in VoIP (Voice over IP) telephony conditions. The speaker verification problem was partly solved over traditional PSTN networks (Public Switched Telephone Network), however nowadays it is also important to assess how specific distortions of VoIP transmission influence the speaker verification scores. As a reference database XM2VTS (Extended multi Modal Verification for Teleservices and Security appliations Data Base) containing English speech (strings of digits) was applied. Three coder degradations (PCM, G.711A and G.723.1) and three network conditions were examined in various configurations to estimate the influence of each, coding and transmission degradation for the final verification scores.