Filtering approaches in spectral domain and features domain have been shown their effectiveness for robust speech recognition. In this paper, we propose a two step filtering method. In the first step, spectral subtraction filter is applied to speech spectrum. In the second step, we design a temporal structure normalization filter in order to apply to features extracted from the filtered spectrum. Our results on Aurora 2 show that the proposed method has higher recognition rate than both of spectral subtraction and temporal filtering methods.