Méthodes avancées pour l’analyse paramétrique non-stationnaire des signaux de la parole et des signaux audio. application à la séparation de source informée

Abstract

Blind Source Separation (BSS) is a topic of great interest in the field of signal processing that has been greatly invested over the last decade. It’s used to solve practical problems of data analysis in various fields of science. BSS is the process of separating a number of signal sources from the mixtures observed from these sources. The most important hypothesis is that the components of the mixtures observed are mutually independent or mutually decorrelated. The BSS extracts the sources by exploiting this hypothesis. The objective of this thesis is to present but also to develop BSS algorithms. These algorithms will then be evaluated by comparing their performances based on a set of blind separation experiments of instantaneous mixtures of audio signal sources. The first part focuses on the Infomax and RobustICA methods applied to audio signals, where we present the simulation results as well as the performance criteria of each method. In the second part of this thesis, the High Particle Swarm Optimization (HEPSO) algorithm which is an improvement of the Particle Swarm Optimization (PSO) algorithm is developed and then used to separate a set of audio signals sources. Compared to the PSO algorithm, the HEPSO algorithm depends on two additional operators. The first operator is based on the multiple crossing mechanism of the genetic algorithm, while the second is based on the mechanism of the bee colony. Both operators were used to update the speed and position of the particles. Thus, they are used to find the optimal separation matrix. The proposed method improves the overall efficiency of the standard PSO method in terms of good exploration and performance. The results of the experiments were done using the BSS demo files