Modeling voice variability through MCE techniques in speaker recognition systems

Speaker recognition is becoming a highly reliable mean for access control and secure-information exchange. However, before its effective use in practical applications, there are still important problems to solve. One of these problems is the degradation of the recognition performance through time due to different factors that introduce a noticeable variability into the voice signal characteristics. In this paper, trying to contribute to the analysis of voice variability in speaker recognition systems, we present some experimental results based on a speech modeling technique known as Gaussian mixture modeling (GMM) trained through a minimum classification error (MCE) criterion. Our major contribution should be to study the vulnerability of the system and to test a start-up process suitable to provide a stable performance along time