Abstract: To solve the recognition dropping problem of speech recognition systems in noise environment , an improved feature extraction algorithm based on perception non-uniform spectral compression is proposed. The algorithm can realize the effective compression of the speech signals in noise environment and make the training and the recognition environments more matching. So the recognition rate can be improved. The experiment result on the intelligent wheelchair platform shows that the algorithm can effectively enhance the robustness of speech recognition, and can ensure the recognition rate in noise environment.