Blind separation of sources that have spatiotemporal variance dependencies

In blind source separation methods, the sources are typically assumed to be independent. Some methods are also able to separate dependent sources by estimating or assuming a parametric model for their dependencies. Here, we propose a method that separates dependent sources without a parametric model of their dependency structure. This is possible by introducing some general assumptions on the structure of the dependencies: the sources are dependent only through their variances (general activity levels), and the variances of the sources have temporal correlations. The method can be called double-blind because of this additional blind aspect: We do not need to estimate (or assume) a parametric model of the dependencies, which is in stark contrast to most previous methods.