We propose a method for separating speech signals when sources outnumber the sensors. In this paper we mainly concentrate on the case of three sources and two sensors. Some existing methods employ binary masks to extract the signals, and therefore, the extracted signals contain loud musical noise. To overcome this problem, we propose the utilization of a directivity pattern based continuous mask, which removes a single source from the observations, and independent component analysis (ICA) to separate the remaining mixtures. Experimental results show that our proposed method can separate signals with little distortion even in a real reverberant environment of T R =130 ms.