In this paper we present an new approach for obtaining an optimal linear transformation of feature vectors. The generalized probabilistic descent method is used in order to optimize the elements of a transformation matrix with respect to a functional approximation of the recognition rate of the training data. The approach is tested in a speaker dependent recognizer for spelled names.
Bibliographic reference. Euler, Stephan (1995): "Integrated optimization of feature transformation for speech recognition", In EUROSPEECH-1995, 109-112.