Separating the individual sources in a single channel is particularly difficult because of overlapping harmonics. Western music is often arranged so that sounds are not only occurring simultaneously, but are also harmonically related. While traditional approaches use either spectral or temporal models, the proposed model exploits the combination of the spectral and temporal correlations of harmonic magnitudes to estimate the regions of overlap. A diverse selection of harmonic musical instruments was analyzed, and a generalized-instrument magnitude track prediction model was derived to generate track estimates. This approach, which exploits dependencies among tracks, was shown to be consistently more accurate than existing methods.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.