Auxiliary Function Based Independent Vector Analysis with Spatial Initialization for Frequency Domain Speech Separation

Abstract

Independent vector analysis (IVA) is one of the state-of-the-art methods for frequency domain speech separation, which can retain the inter-frequency dependency structure to theoretically avoid the classical permutation ambiguity inherent to blind source separation (BSS). Auxiliary function based IVA (AuxIVA) is proposed as a fast form IVA method by adopting the auxiliary function technique to avoid step size tuning. In this paper, the spatial information is introduced as a prior knowledge for AuxIVA to set an initialization, which can not only increase the convergence speed in terms of iteration number but also improve the separation performance. The experimental results with real speech signals and real room recordings confirm the advantage of the proposed method.