【作者】 柳静；

【导师】 赵晓晖；

【作者基本信息】 吉林大学 ， 电子与通信工程， 2012， 硕士

【摘要】 本文主要研究一种基于离散余弦变换的语音增强算法，由于DCT变换相对于DFT变换而言，具有信号能量更为集中、又无需对语音相位进行估计等优点，因此，基于DCT的语音增强算法在保持低运算复杂度的情况下取得了较为理想的语音增强效果。在语音信号处理中，由于语音信号的高度相关性以及短时平稳特性，连续时刻语音DCT分量间一般具有较强的依赖关系。本文对经典的离散余弦变换的语音增强算法进行改进，给出了基于DCT-RLS原理的语音增强算法。该算法利用DCT域中连续时刻语音分量间的相关性，结合递推最小二乘算法的最小均方误差理论，实现纯净语音分量的最优估计。相对于经典的离散余弦变换语音增强算法，本文所提算法的语音增强效果在精度以及收敛速度方面有了新的提高，能有效的提高语音信号的信噪比，抑制语音的失真。更多还原

【Abstract】 Speech enhancement is one of the most important speech processing technology,which is widely used in speech recognition, speech coding and speech synthesis. Thepurpose of the speech enhancement is as far as possible to extract the clean speechsignal from the noisy signal. Because of the diversity of the noise and the complexityof speech signal, the corresponding speech enhancement algorithms are different indifferent noise environment.According to different treatment means of the noisy speech signal, speechenhancement algorithms can be divided into the time domain algorithms andtransform domain algorithms. The algorithms of the time domain directly processspeech signal in the time field, including adaptive denoising algorithm, the algorithmbased on speech production model, etc. The algorithms of the transform domainanalysis and process the speech signal in transform fields, which commonly use DFT,DCT and KLT transform. The representative algorithms include spectral subtractionmethod, sub-space method, the short-term spectral estimation algorithm and thewavelet transformation algorithm, etc.The adaptive algorithm deals with the known input data to seek for an optimalfilter which can minimize the mean square error between the output of the filter andthe needs of the signals. The adaptive algorithms (such as Minimum Mean Squarealgorithm and the Least Square algorithm) usually get the estimation of the filterparameters or the approximation of the optimal filter by recurring the known data. Wecan get the relevant speech enhancement algorithms by using these algorithms inspeech signal field. These algorithms can eliminate the noise of the knownobservation signal effectively. Through computer simulation experiment, we analyze,discuss and get the characteristics and the performance of RLS speech enhancementalgorithm and LMS speech enhancement algorithm. The simulation results show that,RLS speech enhancement algorithm has higher signal-to-noise ratio, and bettercapability in inhibiting signal distortion.The DCT transform has the merits of energy concentration, no need estimatingthe phase of the speech signal. The speech enhancement algorithm based on DCTobtains the ideal effect in keeping low computing complexity. The classical discretecosine transform of speech enhancement algorithm is researched in this paper. Sincethe relevance and short-term stationary characteristic of the speech signal, thecomponents of DCT in continuous time have strong correlation. According to thecorrelation of the components and combining with the theory of minimum mean square error of RLS algorithm, we research on speech enhancement algorithm basedon the DCT-RLS principle and obtain the optimal estimation of clean speech. Thesimulation experiment results show that the algorithm has better performance andhigher Signal-to-Noise ratio of the output speech.更多还原