Aiming at the feature extraction problem in speech emotion recognition,the hierarchical deep belief network (DBN) was used to extract the emotion characteristics of speech signal automatically,and the continuous multi frame speech was combined together to form a high- dimensional abstract feature.The feature trained by deep belief network was used as the input of the limit learning machine (ELM) classifier,and finally a speech emotion recognition system was established.The experimental results in CASIA emotional speech database show that this method is superior to other methods of emotion recognition.