Discriminative Non-negative Matrix FactorizationUsing Interference Factor for Source Separation

Abstract

Non-negative matrix factorization (NMF) is known as efficient basis decompositionmethods for a magnitude spectrogram, which can be used in source separation task.Traditional basis training methods, however, have problem when a trained basis of atarget signal can represent other source signal. In this case, separation performancedegenerated. To resolve this problem, discriminative NMF (DNMF) which trains basisto be discriminative to the other sources have been proposed. In this paper, DNMF byreducing interference factor is proposed. Experiments on speech enhancement was per-formed to show that the proposed method successfully obtained discriminative bases.