Fisher discriminant analysis (FDA) is a popular and powerful method for dimensionality reduction and classification. Unfortunately, the optimality of the dimension reduction provided by FDA is only proved in the homoscedastic case. In addition, FDA is known to have poor performances in the cases of label noise and sparse labeled data. To overcome these limitations, this work proposes a probabilistic framework for FDA which relaxes the homoscedastic assumption on the class covariance matrices and adds a term to explicitly model the non-discriminative information. This allows the proposed method to be robust to label noise and to be used in the semi-supervised context. Experiments on real-world datasets show that the proposed approach works at least as well as FDA in standard situations and outperforms it in the label noise and sparse label cases.