Computational Acceleration And Smart Initialization Of Full-Rank Spatial Covariance Analysis

Full-rank spatial covariance analysis (FCA) is a method for blind source separation. It is based on a model for observation mixtures with flexible source-related parameters, and an EM algorithm is known to optimize the parameters. FCA has the potential to obtain high-quality separations. However, the algorithm for FCA is computationally demanding and sensitive to initializations. This paper proposes two practical techniques to make effective use of FCA. The first one is to accelerate the execution of the algorithm by using single-instruction multiple-data (SIMD) instructions run on a GPU. The second one is to initialize the parameters appropriately by scanning the observation mixtures. Experimental results show that high-quality separations were achieved for 6-second real-room speech mixtures (4 sources and 3 microphones) with a computational time of less than 8 seconds.