Monaural Blind Source Separation in the Context of Vocal Detection

In this paper, we evaluate the usefulness of several monaural blind source separation (BSS) algorithms in the context of vocal detection (VD). BSS is the problem of recovering several sources, given only a mixture. VD is the problem of automatically identifying the parts in a mixed audio signal, where at least one person is singing. We compare the results of three different strategies for utilising the estimated singing voice signals from four state-of-the-art source separation algorithms. In order to assess the performance of those strategies on an internal data set, we use two different feature sets, each fed to two different classifiers. After selecting the most promising approach, the results on two publicly available data sets are presented. In an additional experiment, we use the improved VD for a simple postprocessing technique: For the final estimation of the source signals, we decide to use either silence, or the mixed, or the separated signals, according to the VD. The results of traditionally used BSS evaluation methods suggest that this is useful for both the estimated background signals, as well as for the estimated vocals.