Blind audio source separation of stereo mixtures using Bayesian Non-negative Matrix Factorization

In this paper, a novel approach is proposed for estimating the number of sources and for source separation in convolutive audio stereo mixtures. First, an angular spectrum-based method is applied to count and locate the sources. A nonlinear GCC-PHAT metric is exploited for this purpose. The estimated channel coefficients are then utilized to obtain a primary estimate of the source spectrograms through binary masking. Afterwards, the individual spectrograms are decomposed using a Bayesian NMF approach. This way, the number of components required for modeling each source is inferred based on data. These factors are then utilized as initial values for the EM algorithm which maximizes the joint likelihood of the 2-channel data to extract the individual source signals. It is shown that this initialization scheme can greatly improve the performance of the source separation over random initialization. The experiments are performed on synthetic mixtures of speech and music signals.