Reconstruction de phase par modèles de signaux : application à la séparation de sources audio
Research output: Book/Report › Doctoral thesis › Monograph
|Publication status||Published - Dec 2016|
|Publication type||G4 Doctoral dissertation (monograph)|
This thesis addresses the problem of phase reconstruction in the TF domain for audio source separation. From a preliminary study we highlight the need for novel phase recovery methods. We therefore introduce new phase reconstruction techniques that are based on music signal modeling: our approach consists in exploiting phase information that originates from signal models such as mixtures of sinusoids. Taking those constraints into account enables us to preserve desirable properties such as temporal continuity or transient precision. We integrate these into several mixture models where the mixture phase is exploited; the magnitudes of the sources are either assumed to be known, or jointly estimated in a complex nonnegative matrix factorization framework. Finally we design a phase-dependent probabilistic mixture model that accounts for model-based phase priors.
Those methods are tested on a variety of realistic music signals. They compare favorably or outperform traditional source separation techniques in terms of signal reconstruction quality and computational cost. In particular, we observe a decrease in interferences between the estimated sources and a reduction of artifacts in the low-frequency components, which confirms the benefit of signal model-based phase reconstruction methods.