A SUPERVISED MULTI-CHANNEL SPEECH ENHANCEMENT ALGORITHM BASED ON BAYESIAN NMF MODEL

Abstract

In this paper, we introduce a supervised multi-channel speech enhancement algorithm based on a Bayesian multi-channel non-negative matrix factorization (MNMF) model. In the proposed framework, we consider the probabilistic generative model (PGM) of MNMF, specified by Poisson-distributed latent variables and gamma-distributed priors. In the training stage, the MNMF parameters of the speech and noise sources are estimated via the variational Bayesian expectation-maximization (VBEM) algorithm. In the enhancement stage, the clean speech signal is estimated via the MNMF-based minimum variance distortionless response (MVDR) beamformer. To further improve the enhanced speech quality, we efficiently combine the MNMF-based beamforming technique with a classical unsupervised single-channel enhancement method. Experiments show that the proposed method can provide better enhancement performance than the selected benchmarks.