Single-channel speech separation by including spectral structure information within non-negative matrix factorization

This paper proposes a novel extension on Non-negative Matrix Factorization (NMF) scheme for the separation of single channel speech mixtures, where we impose a post-sparse model on the original weight matrix derived from a previously proposed coherence-constrained NMF model. The approach considers both the modeling ability of NMF basis functions for each source as well as the ability of these basis functions to achieve accurate separation performance. Compared with latest associated NMF models for source separation, the results of our model indicate promising advantages, in terms of both objective source separation measures and Perceptual Evaluation of Speech Quality (PESQ) evaluations.