Abstract: Speech emotion recognition is one of the research hotspots of human-computer interaction. In view of the irrationality of the decision tree in the traditional random forest model (RF), which has the same decision-making power, a differential evolution weighted random forest classification model (DERF) is proposed. In RF, the sample and each node variable are generated randomly, so there will be slight fluctuations in each classification result. In order to improve the system classification stability and recognition accuracy, three identical DERF classifiers are integrated and constructed, and the final decision results are determined according to the majority voting principle. In the experiment, the time domain feature, auditory language spectrum feature and nonlinear hurst parameter feature of speech were combined. The five emotions in the Berlin database and the CASIA database were selected to identify the results. The results show that the proposed method improves the system recognition effectively.