Speech Signal Enhancement Based On Similar Structure

The speech in the reality life inevitably affected by the environment. Speech enhancement is an effective method to solve this problem. Its primary goal is as far as possible to extract pure speech from the noisy speech, and improve the speech quality.In recent years, The signal processing method based on similar structure is obtained attention of scholars.Nonlocal means denoising algorithm as a filtering algorithm using the signal similarity structure, attracted a lot of attention in image and video processing fields. The idea behind the nonlocal means is to exploit the self similarity between two-dimensional patches rather than single pixels. Based on this kind of self similarity,a large number of similar pixels can be collected and averaged, leading to a well denoised signal. Self similarity also exists between one-dimensional fragments of speech signal,and form the basis of the proposed method..According to the features of speech signals, this thesis is the modified nonlocal means(NLM) for speech enhancement under both stationary and non-stationary noise. The main points of this thesis are given as follows:First,this thesis is the modified nonlocal means(NLM) for speech enhancement under both stationary and non-stationary noise. The proposed method consists of two stages:one is the classical spectral subtraction, which takes the role of reducing the amplitude of the noise and some low frequency components in the noise; the other stage is the modified nonlocal means, which collects speech fragments of similar structure and suppresses the background noise in time domain with less musical noise.Second, designing algorithm by nonlocal means in speech power spectrum. First,process the noise speech by the spectral subtraction, and analysis of the residual noise characteristics. A Modified Nonlocal Means is given to eliminate the residual noise.Finally, improve the quality of the speech.