Abstract: In order to reduce the residual noise caused by the method of phase-error based filter, a noise reduction approach based on dual microphones is proposed based on the Gaussian Mixture Model (GMM). First, two cases, i.e. target speech presence and target speech absence were modeled by the GMM, respectively. Then, the frame-based target speech present probability (TSPP) is calculated with the Bayesian classification. Next, an improved PBF (IPBF) is presented by modifying the gain function (GF) of PBF using the rule of noise suppression maximization. Finally, a mask filter for noise reduction is obtained by combining the GF of IPBF and TSPP. Some simulation results reveal that the proposed method could suppress the residual noise effectively, particularly in the period of target speech absence