1.Go over the paper "语音识别中的深度学习方法", which is a thesis of a senior student, graduated from CSLT.
2.Run discriminative training, its result is better than models only use MLE.