Source Separation using the Spectral Flatness Measure

Complex audio scenes with a large number of sound sources pose one of the most difficult problems for audio pattern recognition. Therefore, methods for source separation are very important in this context. Many source separation methods try to exactly recover every source in an audio scene. In this paper, however, we propose an algorithm for the extraction of simpler components from complex audio scenes based on an optimisation approach using a sound complexity measure derived from the spectral flatness measure. We yield good separation for artificial mixtures of three signals with time dependent mixing conditions.