COMPARISON OF VARIOUS MIXTURES OF GAUSSIAN PLDA-MODELS IN THE PROBLEM OF TEXT-INDEPENDENT SPEAKER VERIFICATION
«Speech Technology Center-Innovation», Ltd.; leading scientific researcher
A. Y. Sizov
ITMO University, Department of Speech Information Systems; Student
Abstract. Applicability of unsupervised mixtures of PLDA models with Gaussian priors in a i-vector space for speaker verification is studied. Conditions under which the application is advantageous are analyzed for existing training databases. A mixture of two PLDA models is shown to be more effective than a single PLDA model for a cross-channel task.
Keywords: i-vector, joint factor analysis, PLDA mixture, speaker verification