De-noising of Speech Signal Using Wavelet

Abstract

Abstract In this project the wavelet de-noising method is used to remove the additive white Gaussian noise from noisy speech signals. The idea of wavelet de-noising is to remove the noise by discarding small coefficients of the discrete wavelet transform for the noisy speech signal. These coefficients can be removed by applying some kind of thresholding function which removes any coefficient below a specific threshold value and keep any coefficient above it. Then, the signal reconstructed by applying inverse discrete wavelet transform. To evaluate the performance of such algorithm, some kind of performance measure such as signal to noise ratio ( SNR ) can be applied. Several methods for speech de-noising using wavelets were tested to evaluate their performance. Universal thresholding method is used to threshold the wavelet coefficients. This method uses a fixed threshold for all coefficients, and the threshold selection depends on the statistical variance measurement. Interval dependent thresholding method is also tested to find its performance, here the signal is divided into different interval depends on variance change in it. Then, the threshold value is calculated for each subinterval depends on the noise variance of each interval. Setting all details coefficients in the first scale to zero by assuming that most of the noise power in the first level is tested to evaluate the performance such assumption. Different comparisons are tested such as comparing the performance with different threshold selection rules, comparing the performance with different wavelet families, comparing with other filtering technique. The wiener filtering is compared with wavelet de-noising method.