TreeIVA: Determined blind source separation based on hierarchical sparsity via tree-structured graph

The performance of blind source separation (BSS) has been improved through the design of more sophisticated source models. For instance, independent low-rank matrix analysis (ILRMA) significantly outperforms independent vector analysis (IVA) by modeling a finer structure of the sources. At the same time, simpler methods that merely split the frequency bands without modeling this fine structure have achieved separation performance comparable to ILRMA. This observation raises the following question: Is it possible to achieve higher separation performance without explicitly modeling the fine spectral structure? To address this, we propose TreeIVA, a novel approach that hierarchically splits frequency bands based on an underlying tree-structured graph. Furthermore, we develop two optimization algorithms for TreeIVA utilizing the majorization-minimization and proximal splitting frameworks. Experimental results showed that TreeIVA using a tree-structured graph which reflects the frequency characteristics of the signal exhibits robust performance for speech signals and can outperform ILRMA.