There is a big demand for noise reduction to enhance ASR robustness. A great variety of noise reduction methods have been proposed, but almost none of them can reduce non-stationary noises. The authers have proposed an algorithm that can reduce noises that former methods, such as adaptive beamformers, find diffcult to deal with. In this paper, the authors verify the proposed method as a front-end for ASR by phoneme recognition tests. Its feasibility is also discussed under reverberant environments.