Remixing Stereo Music with Score-Informed Source Separation

Musicians and recording engineers are often interested in manipulating and processing individual instrumental parts within an existing recording to create a remix of the recording. When individual source tracks for a stereo mixture are unavailable, remixing is typically difficult or impossible, since one cannot isolate the individual parts. We describe a method of informed source separation that uses knowledge of the written score and spatial information from an anechoic, stereo mixture to isolate individual sound sources, allowing remixing of stereo mixtures without access to the original source tracks. This method is tested on a corpus of string quartet performances, artificially created using Bach four-part chorale harmonizations and sample violin, viola and cello recordings. System performance is compared in cases where the algorithm has knowledge of the score and those in which it operates blindly. The results show that source separation performance is markedly improved when the algorithm has access to a well-aligned score.