Blind signal separation (BSS) consists of recovering unobserved sources from several observed mixtures; the term “blind” comes from the very weak assumptions made about the mixing and the sources. Most known BSS algorithms are classified into three major categories: nonlinear principal component analysis (PCA), independent component analysis (ICA), and entropy maximization. The book aims to provide an investigation of the most recent theoretical results, applications, and trends in the area of blind signal separation.
The first chapter is a brief presentation of the fundamental concepts and the methodologies currently used to solve specific BSS tasks. The instantaneous mixing model of the independent sources, a series of standard working assumptions concerning the mixing matrix (the sources and noise, respectively), and the relationships of BSS with the concepts of ICA and PCA are briefly explored in the first seven sections of the chapter. The convolutive mixing model, which assumes that the observed signals are linear and convolutive mixtures of a finite set of unknown but statistically independent source signals, is presented in the eighth section. Some of the definitions concerning various aspects of the blind-type models are summarized in the last part of the first chapter.
The second chapter deals with the most significant theoretical aspects of convolutive blind signal separation (CBSS). Following a brief presentation of the most frequently used criteria derived from Shannon’s entropy, Kullback-Leibler divergence, the principle of entropy maximization (INFOMAX principle), and the adaptive learning algorithms based on stochastic gradient and natural gradient are presented in the third section of the chapter. Next, the generalized Gaussian density (GGD) model is presented, in order to derive a suitable family of parametric probability distributions as possible signal models. The fifth section concerns strategies for convolutive mixtures based on second-order statistics and higher-order statistics. The authors point out that different learning algorithms--such as the stochastic, natural gradient updates based on entropy maximization--can be extended by analogy to solve similar separation tasks in case of convolutive mixtures.
The third chapter focuses on frequency-domain algorithms as strategies of CBSS for noise reduction for the hearing impaired. Following a general presentation of the behind-the-ear processors of hearing aid and cochlear implant devices and technical arguments explaining the difficulties of hearing-impaired listeners in understanding speech in a noisy environment, the last part of the chapter discusses the advantages and performance of different noise reduction strategies.
The fourth chapter presents a series of conclusions, comments, and suggestions for further research in the BSS field. The list of references contains the most representative published work in the BSS, PCA, and ICA areas.
The book provides both theoretical guidelines and a series of important applications of the BSS methodologies. It could prove valuable to researchers and students involved in different domains of signal processing.