Discriminative and reconstructive basis training for audio source separation with semi-supervised nonnegative matrix factorization

摘要

This paper addresses an audio source separation problem and proposes a new basis training method for semi-supervised nonnegative matrix factorization (NMF). In a conventional semi-supervised NMF, pretrained spectral bases for a target source can represent other undesired interfering sources, which degrade the separation performance. To solve this problem, we propose the training of two types of supervised bases, discriminative and reconstructive, bases for the target source. In the training stage, the discriminative bases are trained to have unique spectral components of the target source to maximize the discrimination ability from the other sources, whereas the reconstructive bases are trained to represent the complete spectra of the target source. The efficacy of the proposed method is confirmed by performing a semi-supervised music source separation.