【作者】 闫峰；

【导师】 王祁；

【作者基本信息】 哈尔滨工业大学 ， 测试计量技术及仪器， 2008， 硕士

【摘要】 小波分析是上世纪九十年代发展起来的数学分析工具,其优良的时频分析能力使得其在信号处理领域获得了空前的发展,作为小波理论的一个重要分支,近年来小波去噪理论也得到了很大的发展和应用。语音信号去噪是语音信号处理领域中一个重要的组成部分,一般都作为预处理模块存在于系统中。一直以来,人们都在宽带加性噪声的模型上进行研究,提出了各种语音去噪方法,尽管在理论上还没有完全解决语音去噪问题,但是有些方法已经实践被证明是有效的。小波分析由于能同时在时域和频域中对信号进行分析,所以它能有效地实现对语音信号的去噪。本文研究小波在语音信号去噪方面的应用,重点研究小波阈值语音去噪。目前已经提出的小波去噪方法主要有三种,模极大值去噪、空域相关滤波去噪以及小波阈值去噪法。阈值法具有计算量小、去噪效果好的特点,取得了广泛的应用。然而在阈值法中,阈值的选取直接关系到去噪效果的优劣。如果阈值选取过小,那么一部分噪声小波系数将不能被置零,从而在去噪后的信号中保留了部分噪声信息;如果阈值选的偏大,则会将一部分有用信号去掉,使得去噪后的信号丢失信息。本文重点研究了小波阈值法,针对不同的阈值函数的选取、阈值处理方法及小波函数的选取做了研究。针对阈值法中高频信号失真的缺点,我们对小尺度上的小波系数做谱减法预处理,之后以一个小阈值去除剩余噪声,大尺度上仍然利用阈值法处理。经过仿真实验表明,这种处理方法较传统的小波阈值法,保留了更多有用信号,减小了去噪后语音信号的失真。我们将仿真的算法移植到DSP平台中,利用SEED-DEC6713模块,设计实时算法,实现了谱减法、小波阈值法以及我们所提出的方法,验证了仿真效果。更多还原

【Abstract】 Wavelet analysis is a mathematical analysis tools which was developed in the 1990s, because of its excellent time-frequency analysis ability.Wavelet analysis has undergone a unprecedented development in the signal processing field. As an important branch of wavelet analysis, wavelet de-noising theory has also got a great development and application. Voice signal de-noising is an important area of voice signal processing, generally as a pretreatment module exists in the system. Scholars often carry through study on the basis of broad band plus noise, and has brought forward many voice de-noising methods. Although in theory, still not completely solve the problem of voice de-noising, some methods have been proved to be effective in practice. Wavelet analysis can simultaneously analyse signal in time and frequency domain, so it can effectively achieve the voice signal de-noising.We study the application of wavelet in voice signal de-noising, focus on the wavelet shrinkage de-noising. There are three main wavelet de-noising methods at present, they are wavelet shrinkage de-noising, model-max de-noising and spatial selection de-noising. wavelet shrinkage de-noising method has a small amount calculation and good de-noising effect, so it has got wide application. But the selection of de-noising threshold directly related to the de-noising effect. Some wavelet coefficients can’t be set zero when the threshold is undersize and parts of noises are retained; some useful signals will be reduced if the threshold is up-size. These cases may degrade the de-noising effect.This article focuses on the wavelet shrinkage de-noising, study different threshold function, different threshold approach and the selection of mother wavelet. For the shortcoming of high frequency signal distortion in the wavelet shrinkage de-noising, we use spectral subtraction to treat with small scales wavelet coefficients and then remove the residual noise with a small threshold; in big scales, we still use threshold method directly. The simulation shows that this method has a better effect than wavelet shrinkage de-noising, it could reserves more high frequency signal, reduce distortion. We design real time de-noising method, and transplant the above-mentioned algorithms to DSP with SEED-EDC6713 module. We completed spectral subtraction, wavelet shrinkage de-noising and the method this article put forward, verify the simulation results.更多还原