基于变分贝叶斯的语音信号盲源分离算法研究

发布时间：2018-05-07 07:04

本文选题：盲源分离 + 变分贝叶斯　；参考：《兰州交通大学》2017年硕士论文

【摘要】：盲源分离算法因为其在科学研究和工程应用领域的广阔发展前景而受到越来越多的研究人员的注意。近几年盲源分离算法开始在信号处理领域得到广泛应用,且由于良好的处理结果而受到广泛关注。但是传统的独立分量分析方法在对语音信号进行分离的时候存在没有考虑噪声对混合系统的干扰,对观测信号和已知的先验信息没有充分利用导致分离效果不理想,以及对语音信号的内在结构特征没有充分考虑等不足。为了克服以上的不足,使语音分离更符合实际情况,提高其应用价值,本文在研究中将变分贝叶斯独立分量法引入有噪语音分离系统进行分析,并针对语音信号内在所包含的时间结构特性,运用自回归模型对语音信号进行建模,提出了基于AR模型的变分贝叶斯独立分量分析算法。然后运用仿真实验和评价指标分析来验证了算法的效果。本文的主要内容有如下几个方面:首先,简要的介绍了盲源分离的理论的相关知识,主要有相应的原理和数学模型,独立分量分析的几种目标函数和优化算法,以及算法前的预处理方法,并对两种经典的传统独立分量分析算法进行了推导和分析。其次,引入了变分贝叶斯独立分量分析对含噪声的语音混合系统进行分离,从贝叶斯网络和贝叶斯推论入手,充分利用了混合系统的先验信息,为了解决混合系统后验概率计算非常复杂的问题,运用了变分近似的方法完成了整个变分贝叶斯独立分量分析的原理推导,并通过与上文两种经典的独立分量分析的算法进行仿真和评价指标的对比,表明该算法结果更优。最后,针对语音信号内在的时间特性,在上文的基础上提出了基于泛化自回归模型的变分贝叶斯独立分量分析算法。该算法的特点是把源信号具有的时间结构和系统噪声的情况纳入一个框架进行学习,运用泛化自回归模型来近似地建模语音信号所具有的时间结构并给出了完整的理论推导过程。最后用变分贝叶斯学习方法分离出含噪声的语音信号。通过与标准的变分贝叶斯独立分量分析的仿真对比证明了改进后的算法的分离效果有了很大的提高。
[Abstract]:Blind source separation (BSS) algorithms have attracted more and more researchers' attention due to their broad development prospects in scientific research and engineering applications. In recent years, blind source separation algorithms have been widely used in the field of signal processing. However, the traditional independent component analysis (ICA) method does not take into account the interference of the noise to the hybrid system when the speech signal is separated, and the incomplete utilization of the observed signal and the known prior information leads to the unsatisfactory separation effect. And the internal structure of the speech signal is not fully considered and so on. In order to overcome the above shortcomings, make speech separation more in line with the actual situation and improve its application value, this paper introduces the variational Bayesian Independent component method into the noisy speech separation system for analysis. According to the time structure characteristic of speech signal, the autoregressive model is used to model the speech signal, and a variational Bayesian independent component analysis algorithm based on AR model is proposed. Then the simulation experiment and evaluation index analysis are used to verify the effectiveness of the algorithm. The main contents of this paper are as follows: firstly, the related knowledge of blind source separation theory is briefly introduced, including the corresponding principle and mathematical model, several objective functions and optimization algorithms of independent component analysis. Two classical independent component analysis (ICA) algorithms are derived and analyzed. Secondly, variational Bayesian Independent component Analysis (VICA) is introduced to separate the noisy speech mixing system. Based on the Bayesian network and Bayesian inference, the prior information of the hybrid system is fully utilized. In order to solve the complex problem of posteriori probability calculation of hybrid systems, the variational approximation method is used to complete the principle derivation of the whole variational Bayesian independent component analysis. The results of simulation and evaluation are compared with the two classical independent component analysis (ICA) algorithms, and the results show that the proposed algorithm is better. Finally, a variational Bayesian independent component analysis (ICA) algorithm based on generalized autoregressive model is proposed based on the inherent temporal characteristics of speech signals. The characteristic of the algorithm is that the time structure of the source signal and the noise of the system are incorporated into a framework for learning. The generalized autoregressive model is used to approximate the time structure of speech signal and the complete theoretical derivation process is given. Finally, the noisy speech signal is separated by variational Bayesian learning method. The comparison with the standard variational Bayesian Independent component Analysis (ICA) shows that the improved algorithm can improve the separation performance greatly.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【参考文献】