In the application of text-independent speaker recognition, this paper puts forward an improved feature extraction of MFCC parameters to supply the inefficient traditional MFCC. Endpoint detection is added in traditional algorithm to remove silence parts. Gaussian shaped filters are used to replace triangular filter banks to improve the accuracy of speaker identification. Gauss mixed model is for speaker recognition. Experiments show that Gaussian shaped filters gain 9.63% performance improvement while proposed MFCC can significantly improve recognition rate by 11.07%. The result indicates that the new method is an effective feature extraction algorithm.