A FAST EM algorithm for Gaussian model-based source separation

We consider the FASST framework for audio source separation, which models the sources by full-rank spatial covariance matrices and multilevel nonnegative matrix factorization (NMF) spectra. The computational cost of the expectation-maximization (EM) algorithm in [1] greatly increases with the number of channels. We present alternative EM updates using discrete hidden variables which exhibit a smaller cost. We evaluate the results on mixtures of speech and real-world environmental noise taken from our DEMAND database. The proposed algorithm is several orders of magnitude faster and it provides better separation quality for two-channel mixtures in low input signal-to-noise ratio (iSNR) conditions.