Two-stage underdetermined speech source separation using frequency normalization

In this paper, we consider the problem of underdetermined blind source separation for anechoic speech recordings. Existing two-stage methods which first estimate mixing matrix and then separate sources are suitable only for instantaneous mixtures and do not cater for anechoic speech recordings. Other time-frequency (TF) methods based on binary masks are found to have limited performance. We here propose a new two-stage technique which includes frequency normalization to estimate mixing matrix followed by a source separation stage involving denormalization process to estimate frequency dependent mixing matrices. Experimental results are provided to demonstrate the advantage of the proposed method over other methods.