Distributed Speech Separation Based on ComplexGaussian Mixture Model

In this paper, the impact of the spatial covariance matrix (SCM) initialization on the performance of the distributed speech separation was studied under the complex Gaussian mixture model. Based on the conditional independence assumption of the recordings of different nodes, the update of the SCM and the posterior probability corresponding to all received signals could be performed per node. Therefore, the SCM corresponding to each node was proposed to be initialized by the correlation matrix of the steering vector based on direction of arrival (DOA). Meanwhile, a DOA self-clustering method was proposed to find the combination of DOA corresponding to the same speaker from different nodes, which guaranteed that different nodes could cooperate. The proposed distributed speech separation and its initialization method has lower computational complexity than the centralized algorithm and avoids the permutation problem. Experimental results validate the effectiveness of the proposed method.