With the development of speech technology, various spoofed speech has brought a serious challenge to the automatic speaker verification system. The object of this paper is replay attack detection which is the most accessible and can be highly effective. This paper investigates discrimination between the replay speech and genuine speech in each sub-band. For sub-bands with discrimination information, we propose a new filter design approach. Finally, experiments are conducted on the ASV spoof 2017 data set using the algorithm proposed in this paper which demonstrates a 60% relative improvement in term of equal error rate compared with the baseline of ASV spoof 2017.