Nonlocal means estimation of intrinsic mode functions for speech enhancement

Abstract

The main aim of this paper is to introduce a new approach to enhance speech signals by exploring theadvantages of nonlocal means (NLM) estimation and empirical mode decomposition. NLM, a patch-based denoisingmethod, is extensively used for two-dimensional signals like images. However, its use for one-dimensional signals hasbeen attracting more attention recently. The NLM-based approach is quite useful for removing low-frequency noisesbased on nonlocal similarities present among samples of the signal. However, there is an issue of under averaging inthe high-frequency regions. The temporal and spectral characteristics of the speech signal are changing markedly overtime. Thus NLM is conventionally not effective to remove the noise components from the speech signal, unlike imagedenoising. To address this issue, initially, the speech signal is decomposed into oscillatory components called intrinsicmode functions (IMFs) by using a temporal decomposition technique known as the sifting process. Each IMF representssignal information at a certain scale or frequency band. The IMFs do not have abrupt power spectral changes over time.The decomposed IMFs are processed using NLM estimation based on nonlocal similarities for better speech enhancement.The simulation result shows that the proposed method gives better performance in terms of subjective and objectivequality measures. Its performance is evaluated for white, factory, and babble noises at different signal to noise ratios.