Improving independent vector analysis in speech and noise separation tasks

Independent vector analysis (IVA) is an efficient multichannel blind source separation method. However, source models conventionally assumed in IVA present some limitations in case of speech and noise separation tasks. Consequently, it is expected that using better source models that overcome these limitations will improve the source separation performance of IVA. In this work, an extension of IVA is proposed, with a new source model more suitable for speech and noise separation tasks. The proposed extended IVA was evaluated in a speech and noise separation task, where it was proven to improve separation performance over baseline IVA. Furthermore, extended IVA was evaluated with several post-filters, aiming to realize an analogous setup to a multichannel Wiener filter (MWF) system. This kind of setup proved to further increase the separation performance of IVA