Hierarchical priors for more than two sources

Investigate whether a hierarchical prior over more than two sources can provide an effective experimental extension of the Gaussian-mixture separation framework while retaining the exact posterior structure used for two sources.

Background

The paper establishes the geometry of real-gain masking for an arbitrary number of sources, showing that the unconstrained per-source optimal gains partition the mixture automatically. However, restrictions such as nonnegative or bounded gains can break this partition when there are more than two sources. The proposed generative model provides an exact posterior for two sources and can, in principle, be extended through a hierarchical binary decomposition, but the experimental consequences of such a construction are not examined.

The unresolved issue is whether a hierarchical prior over more than two sources is practically effective and how its approximation, source-order dependence, and separation performance behave beyond the two-source experiments reported in the paper.