Exploiting Eigenposteriors for Semi-supervised Training of DNN Acoustic Models with Sequence Discrimination