Abstract: For the music annotation,the amount of unlabeled music data is often much more than the labeled ones such that the training results are usually unsatisfying.Although generation model can be suitable for the smaller training data case to some extent and get higher quality results,it may be inferior to the discriminative model in the case of sufficient training data.By combining the advantages of the generation model and the discriminative model,this paper presents a hybrid discriminative restricted Boltzmann machines.The proposed hybrid model can improve the accuracy of the music annotation tasks.The experiment results show that the hybrid model is much better than the traditional machine learning models.Moreover,it is also better than the single discriminative Boltzmann machines for the case that the amount of training data is small and can attain the similar performance to the discriminative model in the case that the amount of training data is sufficient.Besides,the Dropout method is introduced in this paper to improve the model and prevent the overfitting for the smaller training data.