Enforcing consistency in spectral masks using Markov random fields

Localization-based multichannel source separation algorithms typically operate by clustering or classifying individual time-frequency points based on their spatial characteristics, treating adjacent points as independent observations. The Model-based EM Source Separation and Localization (MESSL) algorithm is one such approach for binaural signals that achieves additional robustness by enforcing consistency in inaural parameters across frequency. This paper incorporates MESSL into a Markov Random Field (MRF) framework in order to addition ally enforce consistency in the assignment of neighboring time-frequency units to sources. Approximate inference in the MRF is performed using loopy belief propagation (LBP), and the same approach can be used to smooth any probabilistic source separation mask. The proposed MESSL-MRF algorithm is tested on binaural mixtures of three sources in reverberant conditions and shows significant improvements over the original MESSL algorithm as measured by both signal-to-distortion ratios as well as a speech intelligibility predictor.