Noise Reduction Of Speech By Wavelet Thresholding

Abstract

The noise reduction of speech performed in wavelet domain is rather seldom used. The Fourier transform performs well upon linear time-invariant signal processing and is appropriate for many applications like stationary signal processing. Nevertheless it cannot deal with many transient characteristics of which speech signals possess. The wavelet expansion gives a more accurate local description and separation of signal characteristics. Wavelets can adjust themselves well to suit the signal and reveal its properties. Besides its time complexity is only O(N). Though there are some theories regarding to the wavelet shrinkage, most of them focus on artificial signal or 2-D image signal processing. The determination of the threshold value for the wavelet shrinkage of speech signals is either unsuitable or lacks theoretical foundation. This paper will find out this value from the view of mean-squared-error perspective, and derive it step by step mathematically. Once the threshold could be set accurately, the noise reduction of speech through wavelet shrinkage could be successfully done and it might not alter the characteristics of the speech or distort the speech itself. Experimental results show that our approach gives encouraging results in noise reduction for speech signals.