Research On Methods For Speech Enhancement

Speech enhancement is a process of filtering the noisy speech signal to improve its quality and performance of speech processing system, and to make speech clearer, more understood and more comfortable. Speech enhancement is an important field of speech processing. It is an effective method to solve noise pollution and improve speech quality. The quality of speech often degrades due to the corruption of noise from surrounding environment in the process of speech communication. The noise not only affects the quality and intelligibility of speech, but also causes hearing fatigue, which hinders normal speech activities. Under the situation, it’s necessary to enhance speech, restrain background noise and improve the quality of speech communication by adopting advanced methods of speech signal process to noisy speech. Therefore, the researches on algorithms of speech enhancement have extensive application meaning.The first chapter introduces the background of this thesis, the purpose and significance of speech enhancement, speech enhancement in the research status at home and abroad. And then the characteristics of speech and noise are introduced. It’s necessary for us to understand the speech enhancement algorithm better. The mathematical model of the speech signal generation and the digital and pretreatment of the speech signal are detailed discussed in the second chapter.The common speech enhancement method and some relatively new theory which contribute to understand and study speech enhancement algorithm deeply are classified introduced. The model parameters of speech signal and noise model parameters and the adjustment parameters of adaptive filter are both depend on the signal section (voice section or noise section) to calculate and determine in speech analysis, filter and enhance of voice. Thus, only have the accurately speech signal endpoint detection, can to correct speech processing. So, we summary the whole voice activity detection method and classification in the third chapter. When the SNR is high, we test the voice activity detection based on short-time energy, and better results is obtained. But when the SNR is low,the method will be discounted, so we also introduces another method based on fractional domain local spectral characteristics of the voice activity detection, through the experiment when the method at low SNR also can get very good results.We tell some simple introduction of the spectral subtraction history at the beginning of the fourth chapter, and the spectral subtraction speech enhancement is utilized broadly because it is simple and easy for the real-time processing. And then the traditional spectral subtraction, the improved traditional spectral subtraction and the spectral subtraction which has been improved in this paper are conducted theoretical derivation and the simulation experiments in MATLAB are made. The experiments find that the traditional spectral subtraction algorithm has certain de-noising function and the improvement of the SNR of general is more than10dB. But after subjective listen we can find "music noise" what is obviously similar to the sound of water. In view of the traditional spectral subtraction speech enhancement with obvious "music noise" defect, put forward a kind of improved spectral subtraction, the method can hold the features of weak components effectively, Through theoretical and experimental analysis shows that to a certain extent, using the method of speech enhancement can reduce "music noise". Although the spectral subtraction improved is better than the traditional spectral subtraction, but the remained "music noise" is boring. So a new improved spectral subtraction is proposed. The theoretical analysis and experimental results show that "music noise" is weakened dramatically and the existence of more surds are guaranteed which improves the intelligibility of the speech. When the input low SNR, still retains a higher output SNR, and the speech distortion is very small.Wavelet transform is a method of signal time-frequency analysis, which possesses a great many virtues. Noise reduction based on wavelet is a superior method to abstract useful signal and lay out noise as well as mutation signal, which has expansive utility value. Wavelet is more and more applied in many field because it Can more exactly analyze the local property of signal. Wavelet theory is a newly developed time-frequency analysis technique and because it can more exactly analyze the local property of signal and is especially of interest for the analysis of non-smtionary signal such as speech, sonar seismic signal, etc. The main idea of wavelet thresholding lies in that when noising signal transforms from time domain to wavelet domain, the signal’s wavelet coefficients will spread to all area of wavelet domain. Although the energy of noise is bigger than the signal, its wavelet coefficients are smaller than the signal’s. So we can use thresholding function to cut off the coefficients of noise and use the rest of coefficients to reconstruct the de-noising signal. The fifth chapter describes the basic theory of wavelet transform, introduced the commonly used wavelet function, analysis of speech signal and noise signal after wavelet decomposition in different characteristics. And the wavelet threshold de-noising is studied excessively. In the wavelet threshold function, to solve the insufficiency of the soft and hard threshold A new method with all improved threshold function is proposed in this paper, and the analysis and derivation are presented. Experimental simulation based on MATLAB results show that a satisfying de-noising result call be obtained by this method compared with traditional methods.