Evaluation of emerging frequency domain convolutive blind source separation algorithms based on real room recordings

Abstract

This paper presents a comparative study of three of the emerging frequency domain convolutive blind source separation (FDCBSS) techniques i.e. convolutive blind separation of non-stationary sources due to Parra and Spence, penalty function-based joint diagonalization approach for convolutive blind separation of nonstationary sources due to Wang et al. and a geometrically constrained multimodal approach for convolutive blind source separation due to Sanei et al. Objective evaluation is performed on the basis of signal to interference ratio (SIR), performance index (PI) and solution to the permutation problem. The results confirm that a multimodal approach is necessary to properly mitigate the permutation in BSS and ultimately to solve the cocktail party problem. In other words, it is to make BSS semiblind by exploiting prior geometrical information, and thereby providing the framework to find robust solutions for more challenging source separation with moving speakers.