A modified spectrogram diffusion algorithm for improved harmonic/percussion separation in music signals

We propose a modification to the spectrogram diffusion algorithm used for harmonic-percussion sound separation (HPSS). The algorithm separates the harmonic portion of a music signal from its percussive portion. The modification fits the framework developed in the spectrogram diffusion method and overcomes some of its shortcomings. Our algorithm uses an auxiliary function minimization approach to derive a simple iterative technique that works on the percussive portion of the signal. We also propose a new method of testing the algorithm using an annotated database. Tests show improvement in harmonic/percussion separation with the proposed modification.