Abstract:
The convolution transfer function-based multi-channel non-negative matrix factorization (CTF-MNMF) has been shown to perform well in blind source separation in highly reverberant environments, but its effectiveness may be limited by the source model. An improved version of the CTF-MNMF is proposed, where the generalized Gaussian distribution (GGD) is used as the source model. The domain parameter is introduced into the NMF and the generalized NMF (GNMF) is utilized to model the non-negative scale factors of the GGD, which enhances the robustness of the source model in capturing signal outliers, and thus improves the accuracy of source estimation. An auxiliary function-based method is used to derive an improved formula for updating the separated matrix and non-negative matrix parameters. Simulation results shows that the proposed algorithm achieves better separation performance than the GGD-ILRMA, WPE-ILRMA, CTF-MNMF algorithms for both speech and music input signals.