Toward Finding Optimal Source Dictionaries for Single Channel Music Source Separation Using Nonnegative Matrix Factorization

Source signal separation from single channel polyphonic music has been attracting conspicuous attention from the audio signal processing community recently. In this paper, we propose a single channel music source separation strategy in an informed setting using nonnegative matrix factorization (NMF). For the first time in the audio source separation context, we adopt the local NMF (LNMF) to distill source dictionaries and experimentally demonstrate the outstanding performance of LNMF over conventional NMF in music source separation. We further extend our work alleviating two fundamental drawbacks of NMF: the randomness of the NMF rank selection which leads to infinitely many possible representations of data and the arbitrary initializations of NMF algorithms which give rise to different source dictionaries for different initial conditions even under a fixed NMF rank. For each music source, using a stability criterion, we identify the stablest NMF rank for which the resulting dictionary is the most reproducible over multiple NMF runs with different initializations. Then we propose to obtain a refined dictionary using clustering approaches over dictionaries obtained for different NMF initializations under the stablest NMF rank. Source separation results which ensure the feasibility of the proposed methods are provided.