Fusion of kernel component analysis methods for speaker recognition

We propose a new kernel component analysis method whose objective criterion is designed to achieve the objectives of both kernel principal component analysis and kernel Fisher discriminant analysis. The basic idea is to maximize the distances between the centers of each sub-cluster and the center of the entire data set in a feature space which has an infinite dimension. The method aims at reducing computational complexity and memory requirements for kernel-based feature enhancement methods. The experimental results on speaker identification tasks show that the proposed methods have higher accuracies in various environments even though the computational complexity and memory requirements are much lower than the other kernel methods such as kernel principal component analysis.