Score informed audio source separation using constrained nonnegative matrix factorization and score synthesis
A musical score provides a great deal of information about a piece of music. In this paper we consider the incorporation of a music score to guide source separation on a single channel recording. We propose a method based on synthesizing lines of music in the score. Dynamic time warping (DTW) is used to to fit the synthesized data to the recording. These are then used as prior distributions in a statistical model of the recorded sound mixture. Probabilistic Latent Component Analysis (PLCA) is then used for the source separation. Preliminary results on a Bach work for string orchestra saw good separation with few artifacts.