Abstract
Independent vector analysis (IVA) is a recently proposed method to solve the permutation problem of frequency domain convolutive blind source separation (FD-CBSS). In IVA the single-variate nonlinear function of the traditional independent component analysis (ICA) is replaced by a multivariate nonlinear function to exploit the higher order dependencies between frequency bins and thereby mitigate the permutation problem. However, IVA can still be affected by a block permutation problem. In this paper, we therefore first analyse the reason for this behavior. Then to address this problem, a robust IVA (RIVA) method is proposed which is based on the phase continuity of the unmixing matrix. The experimental results confirm that the proposed method can successfully solve the block permutation problem of IVA. © 2012 IEEE.