Uma investigação sobre métodos de separação cega de fontes sonoras envolvendo representações não-negativas e diversidade espacial

Abstract

The problem of blind source separation finds many applications across different areas, thus justifying the ever increasing number of works in this topic. This work focuses on studying this problem for sound sources, employing non-negative signals’ representations, while also taking advantage of the spatial diversity induced by the use of multiple channels; this particular feature has recently opened up new research directions regarding the proper modeling of multichannel source separation This work studies two different algorithms: NMF-SCM (sound source separation using non-negative matrix factorization and direction-of-arrival-based spatial covariance model), whose model represents the state of the art, taking in consideration not only the characteristics of the sources but also the enviroment into which they were captured on; and NTF (non-negative tensor factorization), whose simplified model is the multichannel equivalent of NMF (non-negative matrix factorization). During the development of this work both algorithms were implemented. A vectorized and parallelized NMF-SCM implementation is presented; and some improvements are proposed to the NTF algorithm, as well as a method for blind determination of the number of sources in multichannel mixtures.