A comparative study between recent wavelet nonthresholding methods and the well-established spectral subtractive and statistical-model-based algorithms for speech enhancement under real noisy conditions

Abstract

In this paper, a comparative study between current wavelet nonthresholding methods for speech enhancement and the well-established spectral subtractive and statistical-modelbased algorithms is performed. The development and evaluation of speech enhancement methods is essential in many branches of telecommunications and entertainment industry. Two classes of speech enhancement methods encompassing both, DFT and wavelet based methods, are evaluated and faced with the current wavelet nonthresholding schemes. Objective analysis taking into account various real noise environments are presented. Questions about wavelet thresholding performance on real noisy environments are discussed. Objective evaluations considering SNR improvement, correlation among original and enhanced sentences and PESQ scores shown that nonthresholding schemes overcome totally the wavelet thresholding methods in PESQ scores, besides performing equally well in noise suppression. A second objective is the identification, among the considered methods, of the more suitable to certain kind of real noisy conditions.