A Modified Approach of Underdetermined Blind Source Separation in Time-Frequency Domain

This years, some underdetermined blind source separation methods achieve good performance in TF domain, such as the methods of using TF distributions (TFDs). These methods assume that the sources are disjoint or non-disjoint in the TF domain. A non-disjoint approach based on subspace representation requires that the number of sources present at a TF point is strictly less than the number of sensors. Recently, a relaxed condition is proposed. This approach only requires the number of the sources at any TF point does not exceed that of sensors. In this paper, we propose a modified approach based on Moore-Penrose pseudo-inverse matrix to solve the problem which the number of sources exceed that of the sensors. Simulation studies proof good performance for this modified approach.