A new approach to discriminative feature extraction using model transformation

This paper deals with a discriminative feature extraction method aiming to increase the discriminative power of a linear feature transform for speech recognition. The transform is based on the linear discriminant analysis (LDA) aNd is optimized discriminatively through a generalized probabilistic descent (GPD) algorithm employing the minimum classification error (MCE) principle. The utilized GPD/MCE algorithm considers two HMM prototypes only, whereas all prototypes have to be adjusted to the current transformation rule. The new approach which we called "extended linear discriminant analysis with model transformation" (ELDA-MT) takes into consideration the prototypes both in the feature space before transformation and in the lower-dimensional feature space after transformation. Thus, the necessary adjustment can be performed by subjecting the prototypes to the current transformation. Speech recognition experiments with ELDA-MT resulted in a significant reduction of word error rate (WER) of relatively 6.2%.