Underdetermined Blind Source Separation of anechoic speech mixtures in the Time-Frequency domain

This paper focuses on the problem of Under-determined Blind Source Separation (BSS) of anechoic speech mixtures. Our algorithm uses the idea of binary Time-Frequency (TF) mask employed in the Degenerate Unmixing Estimation Technique (DUET), but relaxes the strict sparsity assumption in DUET by allowing the sources to overlap in the TF domain to a certain extent. In particular, the number of active sources at any TF point does not exceed the number of sensors. We use the Unsupervised Robust C-Prototypes (URCP) algorithm to estimate the mixing parameters, and then divide the TF points into disjoint groups and overlapped groups to treat them separately. Experimental results show that the proposed method indicates a substantial increase in the Signal-to-Interference Ratio (SIR) comparing with DUET.