Polyphonic music separation based on the simplified energy splitter

In the past years many approaches have been developed that target the separation of polyphonic music material into independent source signals. Due to lack of information on the original signals it is currently practically impossible to extract the original waveforms from their mixture. Thus all of the approaches target the reconstruction of signals that are at least in some way close to the original. For that purpose common features of harmonic sounds are usually exploited. This paper proposes a system that uses frequency-domain instrument models as prior knowledge for reinserting information needed for the separation. The system provides over 18dB Signal to Distortion Ratio for two simultaneous notes, which slowly degrades as the level of polyphony increases. This makes the approach highly applicable both as a standalone separation tool and the ground of other signal manipulation methods.