带噪混叠语音信号盲分离方法研究

发布时间：2018-03-24 07:11

本文选题：盲源分离　切入点：带噪混叠语音　出处：《北京交通大学》2014年硕士论文

【摘要】：语音是人类传播信息和交流的重要媒介,人们可以在多个讲话者的环境中区分和获取自己感兴趣的语音信号,这是人体内部语音理解机理特有的一种能力。如何通过机器从混合的语音信号中分离出各个源信号,成为语音信号处理领域的一个重要问题。盲源分离(Blind Source Separation,BSS)是混叠语音分离的主要方法之一。盲源分离是指在源信号及其混合方式均未知的情况下,仅根据观测到的若干混合信号恢复源信号的过程。目前的盲源分离基本上都是在无噪环境中进行的,但是实际环境中,语音信号不可避免的会受到各种噪声的影响,因此研究带噪混叠语音分离方法具有重要的理论价值和实际意义。本文对带噪混叠语音信号进行研究,结合盲源分离技术,提出了一种有效的解决带噪混叠语音盲分离的方法。首先消除带噪混叠语音信号中的噪声,提高信号的信噪比,然后再将去噪处理后混叠语音信号进行多个说话人的语音分离；主要在去噪部分和语音分离部分对算法进行改进,论文的主要工作包括：第一,在带噪混叠语音信号的噪声消除方面,提出了一种基于改进噪声估计和幅度补偿的改进的谱减法,该方法在有效去除噪声的同时能极大限度的避免源信号受到损伤,为后续进行的混叠语音信号分离工作奠定基础,可以在很大程度上避免由于源信号受到损伤而影响分离效果。第二,在多个说话人的语音分离方面,提出了结合牛顿下降法和优化快速独立分量分析算法(M-FastICA)的改进算法,解决基于负熵的FastICA算法对随机初始分离矩阵敏感并存在局部最大值的问题,算法在保证分离效果的同时减小了对初始值的敏感度、降低了算法的计算迭代次数；同时根据语音信号的分布特性优化选取分离算法中的非线性函数,以提高算法的精度。最后,可以对分离信号进行再消噪处理,从而进一步提升分离语音信号的质量。仿真实验表明,论文所提算法具有很好的分离效果。从相似系数矩阵和最小均方误差两个指标来看,论文所提算法与原始的FastICA算法相比有着更加出色的分离性能,算法迭代次数也下降了60%,降低了算法的复杂度。
[Abstract]:Speech is an important medium for the dissemination of information and communication among human beings. People can distinguish and acquire voice signals of interest to themselves in the context of multiple speakers. This is a unique ability of the human body's internal speech understanding mechanism. How to separate each source signal from a mixed speech signal through a machine, Blind Source Separation (BSS) is one of the main methods of speech separation. Blind source separation (BSS) means that the source signal and its mixing mode are unknown. The current blind source separation is basically carried out in noise-free environment, but in the actual environment, the speech signal will inevitably be affected by various kinds of noise. Therefore, it is of great theoretical and practical significance to study the noisy aliasing speech separation method. In this paper, we study the noisy aliasing speech signal, combine the blind source separation technology, propose an effective method to solve the blind separation of the noisy aliasing speech. Firstly, the noise in the noisy aliasing speech signal is eliminated, and the signal-to-noise ratio of the signal to noise is improved. Then, the speech separation of multiple speakers is carried out after the process of denoising, and the algorithm is improved mainly in the part of denoising and the part of speech separation. The main work of this paper is as follows:. First, an improved spectral subtraction method based on improved noise estimation and amplitude compensation is proposed for noise cancellation of noisy aliasing speech signals. It can lay a foundation for the subsequent work on the separation of aliasing speech signals, and can largely avoid the influence of source signal damage on the separation effect. Secondly, in the aspect of speech separation of multiple speakers, an improved algorithm combining Newton descent method and optimized fast independent component analysis algorithm (M-FastICA) is proposed. To solve the problem that FastICA algorithm based on negative entropy is sensitive to the random initial separation matrix and has local maximum value, the algorithm not only ensures the separation effect, but also reduces the sensitivity to initial value and reduces the number of iterations. At the same time, according to the distribution characteristics of the speech signal, the nonlinear function of the separation algorithm can be optimized to improve the accuracy of the algorithm. Finally, the separation signal can be de-noised and the quality of the separated speech signal can be further improved. The simulation results show that the proposed algorithm has a good separation effect. From the similarity coefficient matrix and the minimum mean square error, the proposed algorithm has a better separation performance than the original FastICA algorithm. The number of iterations of the algorithm is also reduced by 60%, which reduces the complexity of the algorithm.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【参考文献】