Single channel audio source separation by Clustered NMF

Abstract

This paper proposes to incorporate the perceptual quality criteria into a single-channel audio source decomposition method. Unlike the existing methods, the proposed method applies a perceptually weighted Clustered Non-negative Matrix Factorization (PW-CNMF) on magnitude spectrogram of the mixed signal. CNMF decomposes an audio mixture into an additive parts based representation where the parts usually correspond to individual notes. These parts correspond to the basis vectors and Shifted Non-negative Matrix Factorization (SNMF) is used to cluster these bases into sources. Perceptually weighted CNMF algorithm has been tested for the separation of pitched instruments with improved quality of the separated sources.