Sparse methods for speech separation have become a discussed issue in acoustic signal processing. These sparse methods provide a powerful approach to the separation of several signals in the underdetermined case, i.e., when there are more sources than sensors. In this paper, a two-microphone separation method is presented. The proposed algorithm is based on grouping time-frequency points with similar direction-of-arrival (DOA) using a multi-level thresholding approach. The thresholds are calculated via the maximization of the interclass variance between DOA estimates and allow to identify angular sections, wherein the speakers are located with a strong likelihood. These sections define a set of time-frequency masks that are able to separate several sound sources in realistic scenarios and with little computational cost. Several experiments carried out under different mixing situations are discussed, showing the validity of the proposed approach.