Music signal separation using supervised NMF with all-pole-model-based discriminative basis deformation
Hiroaki NAKAJIMA, Daichi Kitamura, Norihiro Takamune, Shoichi Koyama, Hiroshi Saruwatari, Nobutaka Ono, Yu Takahashi, Kazunobu Kondo · 2016
In this paper, we address the music signal separation problem and propose a new supervised nonnegative matrix factorization (SNMF) algorithm employing the deformation of a spectral supervision basis trained in advance. Conventional SNMF has a problem that the separation accuracy is degraded by a mismatch between the trained basis and the spectrogram of the actual target sound in open data. To reduce the mismatch problem, we propose a new method with two features. First, we introduce a deformation with an all-pole model that is optimized to make the trained basis fit the spectrogram of the target signal, even if the true target component is hidden in the observed mixture. Next, to avoid an excess deformation, we limit the degree of freedom in the deformation by performing discriminative training. Our experimental evaluation reveals that the proposed method outperforms conventional SNMFs.