Choice and adaptation of statistical models for single channel singing voice separation

The problem of singing voice extraction from mono audio recordings, i.e., one microphone separation of voice and music, is studied. The approach is based on a priori probabilistic models for two sources, more precisely on Gaussian Mixture Models (GMM). A method for model adaptation to the characteristics of the mixed sources is developed and a comparative study of different models and estimators is performed. We show that the adaptation of the model of music from the non-vocal parts of songs yields good results in realistic conditions.