One microphone singing voice separation using source-adapted models

In this paper, the problem of one microphone source separation applied to singing voice extraction is studied. A probabilistic approach based on Gaussian mixture models (GMM) of the short time spectra of two sources is used. The question of source model adaptation is investigated in order to improve separation quality. A new adaptation method consisting in a filter adaptation technique via the maximum likelihood linear regression (MLLR) is presented with an associated filter-adapted training phase.