Separating Solo and Orchestra from Monaural Audio Using Mixture Models

We describe a new approach to the "desoloing" problem, in which one tries to isolate the accompanying instruments from a monaural recording of a soloist with accompaniment. Our approach is based on explicit knowledge of the audio in the form of a score match - a correspondence between a symbolic score and the music audio, giving the times of all musical events. We employ the familiar idea of masking the short time Fourier transform to eliminate the solo part. The ideal mask is estimated by fitting a model to the data, whose note-based components are derived from the score match. The parameters for our probabilistic model are estimated using the EM algorithm.