An analysis framework of two-level sampling subspace for speaker verification

Using high-dimensional Joint Factor Analysis (JFA) speaker supervectors for the Fishervoice based subspace analysis suffers high computational complexity problem in the model training process. To address this problem, we propose a two-level sampling subspace framework. For the first level of this framework, partial mean vectors are selected from the JFA speaker supervector to form a low-dimensional feature vector. For the second level, PCA is first applied to perform dimension reduction for the feature vector. Several classifiers are then constructed on a collection of random subspaces generated by randomly sampling the reduced feature space. Finally, all classifiers are fused to obtain the final decision. Experimental results on NIST08 show that the proposed framework improves the performance of JFA and Fishervoice by a relative decrease of 13.8% and 7.2% respectively on EER. The minDCF is reduced to 2.19 by using the new model.