Audio signal separation using supervised NMF with time-variant all-pole-model-based basis deformation

We address a novel nonnegative matrix factorization (NMF) with a new basis deformation method to handle various music sounds. Conventional supervised NMF has a critical problem that a mismatch between bases trained in advance and an actual target sound reduces the accuracy of separation. To solve this problem, we proposed an advanced supervised NMF that applies a single time-invariant filter to the bases for making them fit into the target sound. However, this method suffers from limitations on basis deformation ability, especially for transient instrumental sounds. In this paper, we propose a new time-variant all-pole-model-based basis deformation method. Our proposed deformation method consists of two types of filter that individually deforms attack and sustain parts in one note. Each of the all-pole models can be automatically selected and adapted to the open data via a statistical signal sampling approach. Experimental results show that the proposed method outperforms conventional methods in many types of instrumental sound.