1 Cognitive Systems, Department of Informatics and Mathematical Modeling, Technical University of Denmark2 Department of Informatics and Mathematical Modeling, Technical University of Denmark3 Department of Applied Mathematics and Computer Science, Technical University of Denmark
We recently introduced two algorithms for sparse non-negative matrix factor 2-D deconvolution (SNMF2D) that are useful for single channel source separation and music transcription. We here extend this approach to the analysis of the log-frequency spectrograms of a multichannel recording. The model proposed forms a non-negative tensor factor 2-D deconvolution (NTF2D) based on the parallel factor (PARAFAC) model. Two algorithms are given for NTF2D; one based on least squares the other on Kullback-Leibler divergence minimization. Both algorithms are extended to give sparse decompositions. The algorithms are demonstrated to successfully identify the components of both artificially generated as well as real stereo music.