In speech emotion recognition system, recognition rates will drop drastically when the training and the testing utterances are from different corpora. To solve this problem, a novel sparse feature transfer approach is proposed. By employing sparse coding algorithm, the common sparse feature representation of emotion features from different corpora is obtained. Meanwhile, the maximum mean discrepancy (MMD) algorithm is introduced to measure the distance between different distributions, and is used as the regularization term for the objective function of sparse coding. Finally, the robust sparse features are achieved for recognition. Experimental results show that, compared to traditional methods, the proposed approach can significantly improve the recognition rates for cross databases.