Unified approach for audio source separation with multichannel factorial HMM and DOA mixture model

We deal with the problems of blind source separation, dereverberation, audio event detection and direction-of-arrival (DOA) estimation. We previously proposed a generative model of multichannel signals called the multichannel facto rial hidden Markov model, which allows us to simultaneously solve these problems through a joint optimization problem formulation. In this approach, we modeled the spatial cor relation matrix of each source as a weighted sum of the spatial correlation matrices corresponding to all possible DOAs. However, it became clear through real environment experiments that the estimate of the spatial correlation matrix tended to deviate from the actual correlation matrix since the plane wave assumption does not hold due to reverber ation and noise components. To handle such deviations, we propose introducing a prior distribution over the spatial correlation matrices called the DOA mixture model instead of using the weighted sum model. The experiment showed that the proposed method provided 1.94 [dB] improvement compared with our previous method in terms of the the signal-to-distortion ratios of separated signals.