Numerical Investigation of Weight Parameters For Geometrically Constrained Independent Vector Analysis Using Vectorwise Coordinate Descent or Iterative Source Steering

In this paper, we experimentally investigate the effect of weight parameters in the geometrically constrained independent vector analysis (GC-IVA) based on the auxiliary function approach, where the algorithms derived using the vectorwise coordinate descent (VCD), and iterative source steering (ISS) are referred as to GC-AuxIVA-VCD and GC-AuxIVA-ISS, respectively. GC-IVA aims to achieve both high source separation performance by blind source separation and the capability of directional focusing by beamforming technique. Although the previous studies have shown that separation performance is highly dependent on the parameters that weigh the importance of each geometric constraint, the parameter space having been investigated is limited, where the weight parameters are assumed to be the same for the constraints for both the target and interference signals. Furthermore, the lack of guidance for the parameter tuning process makes applying these algorithms difficult. To improve separation performance, we separately investigate the effect of weight parameters for the constraints for the target and interference signals with numerical experiments. Moreover, we present several tips for weight parameter tuning based on the experimental results, which are necessary to bring GC-IVA one step closer to practical applications. Experimental results showed that separately considering the weights for the constraints effectively improved the source-to-distortions ratio (SDR) and source-to-interferences ratio (SIR).