Kernel based Non-linear Feature Extraction Methods for Speech Recognition

In this paper, we report our recent investigation on the extension of heteroscedastic discriminant analysis and maximum likelihood linear transformation algorithms by taking advantage of the kernel method. The kernel-based heteroscedastic discriminant analysis and kernel-based maximum likelihood linear transformation are formulated. A set of preliminary experimental tests apply the above two techniques to full set digit vs. digit speech classification tasks and reduced sample set 10 isolated digits speech recognition. Comparisons with the existing linear and non-linear feature extraction algorithms such as linear discriminant analysis, kernel based linear discriminant analysis, heteroscedastic discriminant analysis and kernel-based heteroscedastic discriminant analysis are made. Discussions on the effectiveness of the proposed methods are also given