Abstract: To aim at the drawbacks of the conventional subspace microphone array speech enhancement method, some improvements are proposed by using masking properties of human ears. Confidence function is used to determine the noise subspace dimension. In noise subspace, the noise power spectrum is estimated by the conditional likelihood. Then, the masking properties are incorporated into the subspace method to estimate the linear filter. Experiments show that compared with conventional algorithms, the proposed approach suppresses noise more effectively and obtains a significant improvement on objective speech quality measures.