Independent vector analysis with multivariate student's t-distribution source prior for speech separation

摘要

The independent vector analysis algorithm can theoretically 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. A super-Gaussian multivariate Student's t-distribution is adopted as the source prior to model the spectrum of speech signals and to mitigate imprecise variance knowledge as is commonplace in non-stationary signal processing. Moreover, the new multivariate source prior can be interpreted as a joint distribution constructed by a t-copula, which can describe the nonlinear inter-frequency dependency. Experimental results using 50 speech mixtures formed from the TIMIT database confirm the advantages of the proposed algorithm.