Audio motif detection for guided source separation. Application to movie soudtracks.

In audio signal processing, source separation consists in recovering the different audio sources that compose a given observed audio mixture. They are many techniques to estimate these sources and the more information are taken into account about them the more the separation is likely to be successful. One way to incorporate information on sources is the use of a reference signal which will give a first approximation of this source. This thesis aims to explore the theoretical and applied aspects of reference guided source separation. The proposed approach called SPotted REference based Separation (SPORES) explore the particular case where the references are obtained automatically by motif spotting, i.e., by a search of similar content. Such an approach is useful for contents with a certain redundancy or if a large database is be available. Fortunately, the current context often puts us in one of these two situations and finding elsewhere similar motifs is possible. The primary objective of this study is to provide a broad theoretical framework that once established will facilitate the efficient development of processing tools for various audio content. The second objective is the specific use of this approach to the processing of movie soundtracks with application in 5.1 upmixing for instance.