Nonlinear Projection for Cosine Distance Scoring in I-Vector based Speaker Verification Systems.



The i-vector space is nowadays considered as the standard representation of speech information for speaker verification systems. This low dimensional representation of data along with new classifiers such as the Probabilistic Linear Discriminant Analysis (PLDA) or the Cosine Distance (CD) classifier have enabled new progress. The idea of the CD scoring classifier is to map higher dimensional features onto a lower dimensional hypersphere. All current systems first use a classical Linear Discriminant Analysis (LDA) to find the best projection to the lower dimensional space before actually projecting onto the hypersphere. The aim of this work is to define a non linear projection from the higher space directly to the lower dimensional hypersphere which minimizes the intra-class correlation while maximizing the inter-class correlation. The result should be compared to other solutions: Classical CD, PLDA, MultiLayer Perceptron approximation of distance measure and Restricted Bolzmann Machines.