Independent vector analysis for source separation using an energy driven mixed student's T and super Gaussian source prior

Abstract

Independent vector analysis (IVA) can thoretically avoid the permutation problem in frequency domain blind source separation by using a multivariate source prior to retain the dependency between different frequency bins of each source. The performance of the IVA method is however very dependent upon the choice of source prior. Recently, a fixed combination of the original super Gaussian, previously used in the IVA method, and the Student's t distributions has been found to offer performance improvement; but due to the non-stationary nature of speech, this combination should adapt to the statistical properties of the measured speech mixtures. Therefore, in this work we propose a new energy driven mixed multivariate Student's t and super Gaussian source prior for the IVA algorithm. For further performance improvement, the clique based IVA method is used to exploit the strong dependency between neighbouring frequency components. This new algorithm is evaluated on mixtures formed from speech signals from the TIMIT dataset and real room impulse responses and performance improvement is demonstrated over the conventional IVA method with fixed source prior.