Skip to Main content Skip to Navigation

Processus gaussiens pour la séparation de sources et le codage informé

Abstract : Source separation consists in recovering different signals that are only observed through their mixtures. To solve this difficult problem, any available prior information about the sources must be used so as to better identify them among all possible solutions. In this thesis, I propose a general framework, which permits to include a large diversity of prior information into source separation. In this framework, the sources signals are modeled as the outcomes of independent Gaussian processes, which are powerful and general nonparametric Bayesian models. This approach has many advantages: it permits the separation of sources defined on arbitrary input spaces, it permits to take many kinds of prior knowledge into account and also leads to automatic parameters estimation. This theoretical framework is applied to the informed source separation of audio sources. In this setup, a side-information is computed beforehand on the sources themselves during a so-called encoding stage where both sources and mixtures are available. In a subsequent decoding stage, the sources are recovered using this information and the mixtures only. Provided this information can be encoded efficiently, it permits popular applications such as karaoke or active listening using a very small bitrate compared to separate transmission of the sources. It became clear that informed source separation is very akin to a multichannel coding problem. With this in mind, it was straightforwardly cast into information theory as a particular source-coding problem, which permits to derive its optimal performance as rate-distortion functions as well as practical coding algorithms achieving these bounds.
Document type :
Complete list of metadata

Cited literature [202 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Thursday, February 21, 2013 - 11:57:34 AM
Last modification on : Monday, May 17, 2021 - 5:48:03 PM
Long-term archiving on: : Wednesday, May 22, 2013 - 5:02:38 AM


Version validated by the jury (STAR)


  • HAL Id : pastel-00790841, version 1



Antoine Liutkus. Processus gaussiens pour la séparation de sources et le codage informé. Autre. Télécom ParisTech, 2012. Français. ⟨NNT : 2012ENST0069⟩. ⟨pastel-00790841⟩



Record views


Files downloads