Spatially-Regularized Switching Independent Vector Analysis

This paper proposes a novel algorithm that uses spatial regularization and a switching filter to enhance the separation accuracy of Independent Vector Analysis (IVA) with specified source permutation based on prior knowledge of sources' Directions-of-Arrival (DOAs). We call this algorithm Spatially-Regularized Switching Independent Vector Analysis (SR-SwIVA). Switching IVA (SwIVA) is a blind source separation (BSS) algorithm that builds on IVA and improves its separation accuracy with a switching filter. However, SwIVA requires relatively reliable estimates of acoustic transfer functions (ATFs) from sources to microphones for its initialization, which are difficult to obtain from DOAs in real reverberant environments. To overcome this limitation, we introduce spatial regularization to SwIVA, which has been shown to be effective for conventional BSS techniques to align permutation (or order) of separated sources with given DOAs. We conduct simulation experiments to show that our proposed SR-SwIVA can significantly improve the source separation accuracy using ATFs estimated from DOAs while completely eliminating the source permutation alignment errors. The achieved source separation accuracy is comparable to that of SwIVA with oracle ATFs.