Long-Term Reverberation Modeling for Under-Determined Audio Source Separation with Application to Vocal Melody Extraction

In this paper, we present a way to model long-term reverberation effects in under-determined source separation algorithms based on a non-negative decomposition framework. A general model for the sources affected by reverberation is introduced and update rules for the estimation of the parameters are presented. Combined with a wellknown source-filter model for singing voice, an application to the extraction of reverberated vocal tracks from polyphonic music signals is proposed. Finally, an objective evaluation of this application is described. Performance improvements are obtained compared to the same model without reverberation modeling, in particular by significantly reducing the amount of interference between sources.