Modeling prior belief for speaker verification SVM systems

Support vector machines (SVMs) can be interpreted as a maximum a posteriori (MAP) estimation of a model’s parameters, for an appropriately chosen likelihood function. In the standard formulation for SVM classification and regression problems, the prior distribution on the weight vector is implicitly assumed to be a multidimensional Gaussian with zero mean and identity covariance matrix. In this paper we propose to relax the assumption that the covariance matrix is the identity matrix, allowing it to be a more general block diagonal matrix. In speaker verification, this covariance matrix can be estimated from held-out speakers. We show results on two speaker verification systems: a Maximum Likelihood Linear Regression (MLLR)-based system and a prosodic system. In both cases, the proposed prior model leads to more than 10% improvement in equal error rate (EER) with respect to results obtained using the standard prior assumptions.