Abstract: Rhythm of speech is an important acoustic distinction between different Chinese dialects，and the difference feature is an important reflection of rhythm.While difference features ΔMFCC & ΔΔMFCC are used as characteristic parameters and Gaussian Mixture Model（GMM） is used as a trained model，the dialect can be identified through calculating the likelihood probability of the test samples.Changsha dialect，Shaoyang dialect，Hengyang dialect and Mandarin have been investigated with this method，and its effect has been compared with the effect using MFFC as characteristic parameters.Experiment results show that a more high recognition rate and better anti-noise performance can be obtained by GMM trained with difference feature.