Estudo da melhoria da taxa de aprendizagem de um algoritmo para separação cega de fontes utilizando técnicas vindas da teoria de redes neurais

Abstract

ABSTRACT: The main purpose of this work is to study the adoption of some techniques originated in the neural network area to dynamically adapt the step size used during the separation process of convolved mixtures. These techniques are studied in order to analyze their influence on both the convergence and stability of the algorithm adopted in this essay. The techniques adopted on this work for the step size adaptation process are used to achieve the blind separation of real and synthetic convolved mixtures. In addition to the study of the techniques to dynamically adapt the step size, some modifications on the adopted algorithm were evaluated with the purpose of diminishing the computational complexity and improving the convergence presented by it. Furthermore, simulations involving the separation of speech signals convolutively mixtured with noises were also run and evaluated.