Speaker Dependent Threshold's Estimation for GMM-UBM Speaker Verification Systems
Bachir Mehemmel, Mehemmel Abbas, Samir Akrouf · ASME Press eBooks · 2011
In real applications of speaker verification [1] based on GMM-UBM method, the size of training data varies from a speaker to another; the choice of a common threshold of verification will reduce the performance of the system when we only have short records of training data. In this paper we first describe speaker verification using adapted gaussian mixture models based on Reynolds works [2] and the performance obtained by this method. We then propose a new method to estimate a speaker-dependent threshold and the results obtained by this approach.