Nonlinear speech enhancement: An overview

This paper deals with the problem of enhancing the quality of speech signals, which has received growing attention in the last few decades. Many different approaches have been proposed in the literature under various configurations and operating hypotheses. The aim of this paper is to give an overview of the main classes of noise reduction algorithms proposed to-date, focusing on the case of additive independent noise. In this context, we first distinguish between single and multi channel solutions, with the former generally shown to be based on statistical estimation of the involved signals whereas the latter usually employ adaptive procedures (as in the classical adaptive noise cancellation scheme). Within these two general classes, we distinguish between certain subfamilies of algorithms. Subsequently, the impact of nonlinearity on the speech enhancement problem is highlighted: the lack of perfect linearity in related processes and the non-Gaussian nature of the involved signals are shown to have motivated several researchers to propose a range of efficient nonlinear techniques for speech enhancement. Finally, the paper summarizes (in tabular form) for comparative purposes, the general features, list of operating assumptions, the relative advantages and drawbacks, and the various types of nonlinear techniques for each class of speech enhancement strategy.