Simple and powerful instrument model for the source separation of polyphonic music

摘要

This article presents a new approach to sound source separation. The introduced algorithm is based on spectral modeling of real instruments. The separation of independent sources is carried out by dividing the energy of the mixture signal based on these instrument models. This way it is possible to regain some of the information that was lost when the independent sources were mixed together into a single signal. The paper presents the theory behind the proposed separation system, then focuses on the instrument model that is the basic element of the approach. Measurement results are given for polyphony levels from 2 to 10 demonstrating the separation quality, with special regard to the effect of prints on the result.