Abstract: At present, the main problems faced by speech emotion recognition are the lack of consistency in the research results on the relationship between speech acoustic features and emotions, the same characteristics use different databases, the recognition results will vary greatly. Using support vector machine as the recognition machine, feature selection is performed through BP neural network, and the highest recognition rate of EMO-DB database feature combination is 85.59%, the highest recognition rate of the CASIA database feature combination is 74.75%, which improves the speed of the operation. This paper contains two speech databases, one of which is Chinese and one German. After BP neural network feature selection, the recognition rate of the mixed databases experiment of the EMO-DB databases and the CASIA databases was 72.34%.And compared with the articles of the past three years, the experimental results of this paper are at a relatively high level.