Noise-Thresholding with Empirical Mode Decomposition for Low Distortion Speech Enhancement

Degrading the quality and intelligibility of the speech signals, background noise is a severe problem in communication and other speech related systems. In order to get rid of this problem, it is important to enhance the original speech signal mainly through noise reduction. Speech enhancement is the term used to describe such algorithms and devices whose purpose is to improve some perceptual aspects of the speech for the human listener or to improve the speech signal so that it may be better exploited by other speech processing algorithms. Development and widespread deployment of digital communication systems during the last twenty years have brought increased attention to the role of speech enhancement in speech processing problems. For this purpose, this thesis presents novel speech enhancement methods based on applying some thresholding strategies in Empirical Mode Decomposition (EMD) domain. Since speech signals are nonlinear and non-stationary in nature, the performance of related studies is significantly dependent on the analysis method. Although Fourier transform and wavelet analysis made great contributions, they suffer from many shortcomings in the case of nonlinear and non-stationary signals. The EMD, recently been pioneered by Huang et. al. as a new and powerful data analysis method for nonlinear and non-stationary signals has made a novel and effective path for speech enhancement studies. Basically, EMD is a data-adaptive decomposition method with which any complicated data set can be decomposed into zero mean oscillating components, named intrinsic mode functions (IMFs). Such functions give sharp and meaningful identifications of instantaneous frequencies. Recent studies have shown that with EMD, it is possible to successfully identify the noise components from the IMFs of the noisy speech. For instance, in case of white noise, most of the noise components of a noisy speech signal are centered on the first three IMFs due to their frequency characteristics. Thresholding is a widely used process in noise reduction algorithms. The idea is to determine a threshold value and to apply different subtraction algorithms for the segmented regions. However, it is never easy to identify and remove the noise components while keeping the original speech components non-degraded. That is why; one of the major drawbacks of these kinds of processes is the degradation of the speech signal, especially in the process of noisy signals with high signal-to-noise ratios (SNR). In order to minimize the degradation of the original speech components, a modified soft-thresholding strategy that works on a frame basis is adapted in this study. The IMFs of the noisy speech signal are denoised by applying the modified soft-thresholding strategy on the coefficients of each IMF. With the proposed strategy, most of the noise components are successfully removed while the speech components are mainly kept. This strategy enables even signals with high SNRs to be processed effectively. It is never possible to remove all the noise components in a noise reduction method. The remaining noise parts may result in an irritating sound which is referred as the musical noise. That is why; most speech enhancement algorithms not only introduce speech distortion but also suffer from the musical noise artifact. The proposed EMD based algorithm is highly effective in noise removal and introduces a rather discrete noise than a continuous musical sound. In order to obtain even better results for speech quality and quantity, the method was further improved by introducing a Discrete Cosine Transform (DCT) based thresholding as a first stage. The twostage algorithm gives efficient results, successfully improving the SNR of the noisy speech and removing most of the noise components. The thesis work mainly concentrates on white noise; however the method has been further improved with a sub-band approach so that it may be applied to different noise types.