Speech enhancement based on statistical modeling of teager energy operated perceptual wavelet packet coefficients and adaptive thresholding function

Abstract

In order to handle the practical situations of real-life applications, a speech enhance- ment method is needed to be capable of producing optimum results with improved overall speech quality with maximized intelligibility particularly under low levels of SNRs. For solving this open problem, this thesis presents a speech enhancement approach, where an adaptive threshold is statistically determined using the Tea- ger energy (TE) operated perceptual wavelet packet (PWP) coefficients of noisy speech. A frame of noisy speech signal is analyzed first in PWP transform domain to obtain a set of PWP coefficients. TE operation is performed on the PWP coeffi- cients to increase the separability between clean speech and noise coefficients. The TE operated PWP coefficients with better time and frequency resolution are then used to determine an appropriate adaptive threshold based on different statistical models, namely Gaussian, Laplace, Rayleigh, Poisson and Student t distributions. The threshold thus obtained is applied upon the PWP coefficients by employing a custom thresholding function, which is designed based on the presence of noise in the noisy speech signal. A couple of custom thresholding functions designed in this thesis can be viewed as a linear combination of the modified hard or μ-law thresh- olding function and the semisoft thresholding function. The enhanced speech frame is synthesized by performing the inverse PWP transform on the thresholded PWP coefficients obtained using the statistically determined threshold and the designed custom thresholding function. The final enhanced speech signal is reconstructed by using the standard overlap-and-add method. Extensive Simulations using NOIZEUS database are carried out considering the presence of car and multi-talker babble noises to evaluate the performance of the proposed method in terms of standard ob- jective metrics and subjective listening tests. It is shown that the proposed method outperforms the reported state-of the-art methods with superior efficacy at high as well as low levels of SNRs.