In order to solve the low dialect identification rate because PCA doesn’t effectively use the sample classification information, a method of feature extraction using PCA and LDA is employed. In this paper, PCA is used to effectively reduce the dimensions of Mandarin, Shanghainese, Cantonese, Minnanese, and then LDA is adopted to extract feature vectors from the dimension-reduced space as the input vectors with BP neural network to recognize. The Simulation results demonstrate that the average dialect identification rate based on PCA and LDA can be up to 85%.