Directional NMF for joint source localization and separation

We propose an unsupervised method for simultaneously localizing and separating speech signals by factorizing a non-negative matrix of Steered Response Power (SRP) measurements. We use a probabilistic interpretation of the SRP function to compute a wideband SRP matrix. Non-negative Matrix Factorization (NMF) is used to decompose it into three terms that describe (1) the source distributions over spatial directions, (2) the overall source activations, and (3) the source activations over the time-frequency (TF) plane. The first term indicates the sources' Directions-of-Arrival (DOA) and the latter two terms provide TF weights for separating the sources. Experiments show that this joint approach out-performs a sequential SRP localization + beamforming method.