This paper describes how kernel methods for discriminant analysis can be applied to speech recognition. The standard Linear Discriminant Analysis (LDA) is used for reduction of feature vector components in the feature extraction. The Kernel Discriminant Analysis (KDA) is non-linear expansion of this technique to project the feature vectors onto the best discriminantive feature, while the non-linear projection is implicity performed by the so called kernel trick. This is a way to represent the scalar-product of non linearly transformed feature vectors without performing the transformation itself. The resulting formulation is expressed as an eigenvalue problem, similar to the linear one. The main difficulty is that the size of the eigenvalue problem is equal to the number of input training vectors, which can be very large in speech recognition. To get the largest eigenvalue and the corresponding eigenvectors we use the efficient Lanczos Algorithm. Preliminary results using the Kernel Discriminant Analysis are presented on a small subset of the resource management RM1 dataset.
Cite as: Katz, M., Krüger, S.E., Schafföner, M., Andelic, E., Wendemuth, A. (2004) Kernel methods for discriminant analysis in speech recognition. Proc. 9th Conference on Speech and Computer (SPECOM 2004), 95-98
@inproceedings{katz04_specom,
author={M. Katz and S. E. Krüger and M. Schafföner and E. Andelic and Andreas Wendemuth},
title={{Kernel methods for discriminant analysis in speech recognition}},
year=2004,
booktitle={Proc. 9th Conference on Speech and Computer (SPECOM 2004)},
pages={95--98}
}