The invention belongs to the technical field of emotion recognition of speech, relates to a cross-corpus emotion recognition method based on adaptive semi-supervised non-negative matrix decomposition,and solves the problem that existing cross-corpus speech emotion recognition methods have low accuracy of emotion recognition due to mismatching of a user's speech. As a feature space and a tag spaceshare coding information, speech emotion feature representations which are more discriminative can be learned; differences of the feature representations can be measured via maximum mean difference,so that mismatching of a speaker's speech is lessened; information of the original feature space is retained in new feature representations as far as possible according to the hypothesis of manifold match-up, so that the speech emotion feature representations are more robust. Unweighted mean recall of four groups of cross-corpus speech emotion recognition tasks reaches 43.74%, while weighted meanrecall reaches 43.84%; accuracy of cross-corpus speech emotion recognition is increased effectively. The method of the invention is applicable to the technical field of emotion recognition for speech.