Spatially Regularized Switching Independent Vector Analysis with Iterative Source Steering

Abstract

We present a computationally efficient approach to speech separation based on spatially regularized switching independent vector analysis (SR-SwIVA) incorporating an iterative source steering (ISS) update. By introducing a switching mechanism, the proposed framework models the timevarying characteristics of multichannel mixtures using multiple separation matrices, while spatial regularization based on direction-of-arrival information solves the inter-state permutation problem. To further improve efficiency, the conventional matrix-inversion-based update is replaced with an ISSbased rank-one update, significantly reducing computational cost while preserving separation performance. Experimental results in noisy and reverberant environment demonstrate that the proposed method achieves improved separation quality and faster convergence compared with conventional SRSwIVA, making it suitable for practical scenarios with limited microphone arrays.