An Adaptive Speech Signal De-noising Algorithm Based On Estimation Of Scaled Noise Energy And Its Implementation

Speech related technologies have been applied to a growing number of fields as the development of information and technology. Due to the great complexity of the practical environments, results of the speech related applications are far from satisfaction with the interference of background noise, which indirectly hampers the speed of further developing speech technologies. So, the researches of speech pre-processing are with great theoretical and practical signification.Speech signal is a time-variable and unstable signal, so it'd be a challenge for the traditional signal processing methods such as Fourier transform, to deal with signals like speech. The wavelet transform is an emerging theory of signal processing, which was developed since late 80s. Unlike traditional signal processing theories, wavelet transform has a better resolution on both time and frequency fields for signal. Therefore, it has a better performance to analyze the unstable, time-variable signals.This paper introduces the basic theory of wavelet transform, from its theoretical background to its applications. Later, we discuss the routine of wavelet threshold de-noising scheme. The critical steps such as the choice of wavelet basis and of the optimal decomposition level, the construction of thresholding function and the estimation of threshold are further studied. Based on the wavelet threshold de-noising theory, a novel scheme of wavelet thresholding for speech signal de-noising is proposed. The scheme introduces a new thresholding function to overcome the drawbacks of traditional soft and hard thresholding functions. This newly introduced function is continuous at the threshold, so that it can curb the Pseudo-Gibbs phenomena effectively. Besides, it can reduce constant bias between the original wavelet coefficient and the estimated one, which helps to preserve the feature of the signal. With this threshold function, an adaptively selective algorithm of wavelet threshold is then proposed. This algorithm has both the virtues of Spatial Selective Noise Filtration method and the traditional threshold de-noising method. It takes advantage of the inter-scale correlations of the wavelet coefficients to estimate the energy of scaled noise signal. In this algorithm, the threshold in each scale is selected based on the scaled noise energy. The simulations demonstrate that the proposed method is superior to some other existing adaptive wavelet de-noising methods. Under such a theoretical background, this paper then presents a speech processing software called Speech Pre-processing Kit, within which also demonstrates the achievements we have in the research on wavelet de-noising theory.