Method and system for generating advanced feature discrimination vectors for use in speech recognition