Abstract: Speaker recognition system is susceptible to noise.On this problem,the non-stationary of the speech signal needs to be considered based on the assumption that the energy and information are different in different sub-band signals.So,the stability of the noise and the completeness of the information for different sub-band signals are analyzed.Through a large number of experiments,it can be found that the low frequency speech signal,which can not only bring a large number of personalized information but also concentrate much energy,is not easy being interfered by noise.Some appropriate treatment has been adopted to make the system's recognition rate over 90% for the signal to noise ratio of 10dB.