Because MFCC can't reflect the dynamic characteristics of speech signal and their own non-stationary, a feature extraction method by combining time-frequency distribution with MFCC is proposed. First get time-frequency distribution of speech signal, and convert time-frequency domain into frequency domain, then extract MFCC+MFCC as characteristic parameters. Finally speaker recognition uses the support vector machine. The simulation experiment compares recognition performance when MFCC and MFCC+MFCC are respectively as characteristic parameters by speech signal and all kinds of time-frequency distribution. Results show that the speaker recognition performance using MFCC+MFCC based on the CWD time-frequency distribution can be improved to 95.7%.