Informed Source Extraction based on Independent Vector Analysis using Eigenvalue Decomposition

A desired acoustic source can very often only be observed in a mixture together with interfering sources in a real-life scenario. Hence, extracting the desired signal with a minimum amount of information about the geometric and acoustic setup is a problem of great interest. Recently, methods for blind source extraction based on Independent Vector Analysis (IVA) have been proposed. These algorithms are entirely blind, which prevents them from focussing on a specific source in the mixture. In this contribution, we guide the convergence of the extraction filter by a free-field prior within a Bayesian model towards the desired solution and use recently proposed update rules relying on the Eigenvalue Decomposition (EVD) for its optimization. The superiority of the presented update rules over a recently proposed state-of-the-art method is shown in experiments using measured Room Impulse Responses (RIRs).