Online Score-Informed Source Separation with Adaptive Instrument Models

In this paper, an online score-informed source separation system is proposed under the Non-negative Matrix Factorization (NMF) framework, using parametric instrument models. Each instrument is modelled using a multi-excitation source-filter model, which provides the flexibility to model different instruments. The instrument models are initially learned on training excerpts of the same kinds of instruments, and are then adapted, during the separation, to the specific instruments used in the audio being separated. The model adaptation method needs to access the musical score content for each instrument, which is provided by an online audio-score alignment method. Source separation is improved by adapting the instrument models using score alignment. Experiments are performed to evaluate the proposed system and its individual components. Results show that it outperforms a state-of-the-art comparison method.