A new noise-robust speecb tecognition method is presented based on the difference in the power spectrum. The idea is to remove tile additive noise by filtering in the power spectrum domain. Feature extraction is carried out in two steps: (i) tbe sbort-tilne power spectrum of the speech signal is allowed to pass through a filter bank; and (ii) the differences in the filter outputs are calculated. Theoretical analysis and experimental results show tbat using the proposed featurcs can significantly improve the recogniser's performance in a noisy environment.