语音信号的盲分离技术研究及应用
发布时间：2018-09-05 14:06
【摘要】：语音信号的盲分离指的是在语音源信号的信息未知,同时也不知道混合系统的情况下,只能依据观测到的语音信号来估计源信号的过程。主要用于现代的通信领域。随着人类步入信息社会的步伐加快,越来越多的领域需要语音信号盲分离。语音信号盲分离研究主要在时域和频域上展开。时域内只能解决瞬时混合模型的语音信号,而有混响问题的卷积语音信号盲分离在频域内解决就容易很多。目前,语音盲分离领域已经出现了很多算法,主要分为批处理算法和自适应算法两大类。批处理算法主要是联合对角化算法,自适应算法则以在线学习的梯度算法为主,有随机梯度算法和自然梯度算法。大量的学者在这两类算法上探索研究改进,Fast ICA算法应运而生,结合了批处理算法和自适应算法的优点,对接收到的数据实施在线梯度算法,一次处理批量的数据并不断更新迭代分离矩阵,有很好的收敛性能。跟其它的ICA算法一样,FastICA算法也存在排序不确定性和幅度不确定性,本文对这部分难点做了大量仿真对比研究,最后实现了频点对齐并消除了幅度不确定性。本文主要做了以下方面的研究:1.分析比较了ICA的四类独立性判断准则,极大化似然度的判断准则、互信息最小化判断准则、信息最大化准则以及极大化非高斯性准则;研究分析了ICA中传统的批处理算法及自适应处理算法。2.深入讨论FastICA算法,并根据目标函数的不同,分别基于负熵、基于峭度以及基于似然度,分析比较三者的优缺点。3.针对排序不确定性和幅度不确定性做了大量对比研究,最终通过基于功率比的相关系数和最小失真法消除了这两种不确定性,并给出了仿真结果和对比分析。4.在语音盲分离的应用方面,探究其在麦克风阵列中的实际应用,分别对实际的欠定语音信号和超定语音信号成功的实现盲分离。
[Abstract]:The blind separation of speech signals refers to the process of estimating the source signals only based on the observed speech signals when the information of the speech source signals is unknown and the mixed system is not known at the same time. Mainly used in modern communication field. With the acceleration of human step into the information society, more and more fields need blind separation of speech signals. Blind speech signal separation is mainly carried out in time domain and frequency domain. In time domain, only the speech signal of instantaneous mixing model can be solved, but the blind separation of convolutional speech signal with reverberation problem is much easier in frequency domain. At present, there are many algorithms in the field of speech blind separation, mainly divided into two categories: batch processing algorithm and adaptive algorithm. Batch algorithm is mainly a joint diagonalization algorithm, and adaptive algorithm is an online learning gradient algorithm, with random gradient algorithm and natural gradient algorithm. A large number of scholars have explored and studied the improved Fast ICA algorithm in these two kinds of algorithms. Combining the advantages of batch processing algorithm and adaptive algorithm, the online gradient algorithm is applied to the received data. It has good convergence performance by processing batch data and updating iterative separation matrix. Similar to other ICA algorithms, there are ordering uncertainties and amplitude uncertainties in FastICA algorithm. In this paper, a large number of simulation studies on these difficulties have been done. Finally, frequency alignment has been realized and amplitude uncertainty has been eliminated. This paper mainly does the following research: 1. Four kinds of independence criteria of ICA, maximum likelihood criterion, mutual information minimization criterion, information maximization criterion and maximization non-Gao Si criterion are analyzed and compared. The traditional batch processing algorithm and adaptive processing algorithm. 2. 2 in ICA are studied and analyzed. The FastICA algorithm is discussed in depth, and the advantages and disadvantages of the three algorithms are analyzed and compared based on negative entropy, kurtosis and likelihood respectively according to the difference of objective function. A large number of comparative studies have been done on sequencing uncertainty and amplitude uncertainty. Finally, the two uncertainties are eliminated by correlation coefficient and minimum distortion method based on power ratio, and the simulation results and comparative analysis are given. In the application of speech blind separation, the practical application in microphone array is explored, and the blind separation of actual underdetermined speech signal and overdetermined speech signal is realized successfully.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3
本文编号：2224504
[Abstract]:The blind separation of speech signals refers to the process of estimating the source signals only based on the observed speech signals when the information of the speech source signals is unknown and the mixed system is not known at the same time. Mainly used in modern communication field. With the acceleration of human step into the information society, more and more fields need blind separation of speech signals. Blind speech signal separation is mainly carried out in time domain and frequency domain. In time domain, only the speech signal of instantaneous mixing model can be solved, but the blind separation of convolutional speech signal with reverberation problem is much easier in frequency domain. At present, there are many algorithms in the field of speech blind separation, mainly divided into two categories: batch processing algorithm and adaptive algorithm. Batch algorithm is mainly a joint diagonalization algorithm, and adaptive algorithm is an online learning gradient algorithm, with random gradient algorithm and natural gradient algorithm. A large number of scholars have explored and studied the improved Fast ICA algorithm in these two kinds of algorithms. Combining the advantages of batch processing algorithm and adaptive algorithm, the online gradient algorithm is applied to the received data. It has good convergence performance by processing batch data and updating iterative separation matrix. Similar to other ICA algorithms, there are ordering uncertainties and amplitude uncertainties in FastICA algorithm. In this paper, a large number of simulation studies on these difficulties have been done. Finally, frequency alignment has been realized and amplitude uncertainty has been eliminated. This paper mainly does the following research: 1. Four kinds of independence criteria of ICA, maximum likelihood criterion, mutual information minimization criterion, information maximization criterion and maximization non-Gao Si criterion are analyzed and compared. The traditional batch processing algorithm and adaptive processing algorithm. 2. 2 in ICA are studied and analyzed. The FastICA algorithm is discussed in depth, and the advantages and disadvantages of the three algorithms are analyzed and compared based on negative entropy, kurtosis and likelihood respectively according to the difference of objective function. A large number of comparative studies have been done on sequencing uncertainty and amplitude uncertainty. Finally, the two uncertainties are eliminated by correlation coefficient and minimum distortion method based on power ratio, and the simulation results and comparative analysis are given. In the application of speech blind separation, the practical application in microphone array is explored, and the blind separation of actual underdetermined speech signal and overdetermined speech signal is realized successfully.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3
【参考文献】
相关期刊论文 前1条
1 孟哲;基于小波变换的多尺度多阈值语音增强方法[J];武汉理工大学学报(交通科学与工程版);2001年02期
，本文编号：2224504
本文链接：https://www.wllwen.com/kejilunwen/wltx/2224504.html
