Single channel blind source separation based on probabilistic matrix factorisation

Abstract

A novel single channel blind source separation method based on probabilistic matrix factorisation (PMF) is proposed. Compared to the conventional non-negative matrix factorisation (NMF) employing Euclidean distance or Kullback-Leibler divergence, PMF uses the log posterior probability as a cost function for optimising spectrum and activation matrices. Such cost function has an advantage that the hyperparameters are optimised numerically without cross-validation. In order to apply PMF to audio source separation, both Gaussian and Laplacian priors are considered. Exponential substitution for target matrices is also proposed to guarantee the non-negativity of the separated spectrogram. In source separation experiments, the proposed PMF-based approach provided significantly better performance than the conventional NMF.