A novel and noise robust front-end based on the combination of spectral noise reduction and Probability Modelbased feature compensation and Cepstral Mean Subtraction (CMS) is proposed. Mel filter-bank outputs can be affected by additive noise primarily because of the vulnerable spectral valleys. An instantaneous Wiener filter is used to improve SNR of the spectral valley. Because the compensated MFCC is an approximation of the clean one and retains a residual mismatch, features are further processed by CMS in order to remove the global shift of the mean. In the presence of additive noise, ASR experiments reveal that a cascade fashion use of these techniques improves recognition performance greatly. For the 863 continuous Chinese speech databases, the average recognition rate across different noise types is improved from 34.63% (using unmodified MFCCs) to 84.63% (using the proposed techniques) at best.