In order to improve the accuracy of emotion recognition, this paper proposes a novel speech emotion recognition algorithm, and takes speech signal as the research subject. The idea of locality preserving projection (LPP) is integrated into the objective function of PCA. It not only realizes the maximum of the total variance, but also keeps the local neighbor structure. That is beneficial to comprehensive extraction of global and local features, and overcomes the defects that the traditional PCA can only keep the structure in global and can not maintain the structure in local. Experiment results verify the feasibility and effectiveness of the proposed method, and accomplish recognition for five kinds of human emotion (joy, anger, sadness, fear, neutral). Research results provide new methods into emotion recognition, promote the further development of human-computer interaction system.