Abstract: Noises cause feature distortion of speech and make the performance of speech recognition system seriously poor. Comparing with classical methods, the histogram equalization can reduce non-linear distortion and improve the robustness of speech recognition system quite well. However in many applications, the feature distribution between training and test speech is usually not identical because of their difference in phonetics or acoustics, then the validity of HEQ can be weaken. The proposed algorithm in this paper utilizes K-means clustering to classify the pre-equalized noisy features into several classes, then further equalizes the features belong to the same class. The experiments show the proposed method improves the performance of system with comparison of usual histogram equalization.