Reduced-rank modeling of time-varying spectral patterns for supervised source separation

In this paper, we propose a new modeling technique of signals having time-varying spectral patterns for supervised source separation. Typical examples of such signals are instrumental sounds having several segments such as “attack” and “sustain”. In the proposed technique, a given signal is modeled as a linear combination of multiple bases which are obtained by using reduced-rank representation of the given signal, where the number of bases is determined automatically. The proposed technique is used to generate the basis matrix in the context of supervised source separation, which improves conventional source separation methods.