In order to improve the reconstruction performance of deep models, reconstruction error constraint based on cross entropy is added to traditional contrastive divergence (CD) algorithm. The improved algorithm is used to train reconstructive deep auto encoder(RDAE), which is used to replace the vector quantization method for LSF in MELP speech coding algorithm. Experimental results show that the improved CD algorithm improves the deep model gain reconstruction performance while costing some likelihood of the model. When the node number of the hidden layer of RDAE is set to 19 bit, the indicators, which include the weighted LSF distance, the performance of reconstructed speech, and the spectrum distortion, perform better in both training set and testing set by the proposed method than by the vector quantization method at 25 bit. That is to say, the coding bitrate of the MELP coder is reduced from 2.5 kb/s to 2.1 kb/s. The reduction rate of the coding bitrate is up to 12.5%, while the speech quality remains.