A low-complexity joint optimization of blind source separation and dereverberation

This paper proposes a low-complexity weighted-prediction-error (WPE) based independent low-rank matrix analysis (ILRMA). Instead of taking the prediction matrix as a whole in WPE-ILRMA, the prediction matrix is expanded to derive the cost function. The minimization of the cost function is simplified using the orthogonality between the mixing filter and demixing filter of different sources, which enables to dereverberate the observed signals with a low complexity. Therefore, the proposed method requires a smaller dimension matrix inverse by exploiting the relationship between the prediction matrix and demixing filter, and has a lower computational complexity than WPE-ILMRA. The cost function is formulated using the maximum log-likelihood criterion, which is then minimized using the coordinate descent method. Experimental results show that the proposed method can achieve a similar separation performance as WPE-ILRMA with lower computational complexity and higher stability.