Conditioned Source Separation for Musical Instrument Performances

In music source separation, the number of sources may vary for each piece and some belong to same family instruments, thus sharing timbral characteristics making more correlated. This leads additional challenges in separation problem. paper proposes a method multiple musical instruments sounding simultaneously explores how much information apart from audio stream can lift quality separation. We explore conditioning techniques at different levels primary network utilize two extra modalities data, namely presence or absence mixture, corresponding video data.