The performance of the speaker recognition system degrades drastically in the noisy environment. A robust feature extraction method for speaker recognition is proposed in this paper. Warped filter banks(WFBS) are used to simulate the human auditory characteristics. The cubic root compression method, relative spectral filtering technique(RASTA) and the cepstral mean and variance normalization algorithm(CMVN) are introduced into the robust feature extraction. Subsequently, simulation experiment is conducted based on Gaussian mixes model(GMM). The experimental results indicate that the proposed feature has better robustness and recognition performance than the mel cepstral coefficients(MFCC) and cochlear filter cepstral coefficients(CFCC).