Linear feature transformation technique is widely used to improve feature discriminability. It can reduce the dimensionality of the feature space, un-correlate the feature components, hence more discriminative model can be obtained. In this paper we compare three discriminative linear transformation approaches in Mandarin digit string recognition (MDSR) system. Compared with the conventional Linear Discriminant Analysis (LDA), two other discriminative linear transformation methods derived from LDA, that is Confusion Discriminant Analysis (CDA) and Heteroscedastic Discriminant Analysis (HDA), are studied on the basis of state-specific confusable class definition and its class-dependent linear transformations.