Generalized Fast Multichannel Nonnegative Matrix Factorization Based on Gaussian Scale Mixtures for Blind Source Separation

This paper describes heavy-tailed extensions of a state-of-the-art versatile blind source separation method called fast multichannel nonnegative matrix factorization (FastMNMF) from unified point view. The common way deriving such an extension is to replace the multivariate complex Gaussian distribution in likelihood function with its generalization, e.g., Student's t and leptokurtic generalized distributions, tailor-make corresponding parameter optimization algorithm. Using wider class distributions scale mixture (GSM), i.e., whose variances are perturbed by positive random scalars impulse variables, we propose GSM-FastMNMF develop expectationmaximization algorithm that works even when probability density variables have no analytical expressions. We show existing FastMNMF instances derive new instance based on hyperbolic include normal-inverse Gaussian, t, as special cases. Our experiments normalinverse outperforms ILRMA model speech enhancement terms signal-to-distortion ratio.