Independent vector analysis with a generalized multivariate Gaussian source prior for frequency domain blind source separation

摘要

Independent vector analysis (IVA) is designed to retain the dependency within individual source vectors, while removing the dependency between different source vectors. It can theoretically avoid the permutation problem inherent to independent component analysis (ICA). The dependency in each source vector is retained by adopting a multivariate source prior instead of a univariate source prior. In this paper, a multivariate generalized Gaussian distribution is adopted as the source prior which can exploit frequency domain energy correlation within each source vector. As such, it can utilize more information describing the dependency structure and provide improved source separation performance. This proposed source prior is suitable for the whole family of IVA algorithms and found to be more robust in applications where non-stationary signals are separated than the one preferred by Lee. Experimental results on real speech signals confirm the advantage of adopting the proposed source prior on three types of IVA algorithm.