盲声源分离技术应用研究

发布时间：2018-05-18 08:16

本文选题：盲源分离 + 线性混合　；参考：《电子科技大学》2014年硕士论文

【摘要】：我们生活在声音的世界里。在嘈杂的环境下,我们很难获得理想的声源,并且交谈会变得不容易。因此从带有噪声的观测信号中获得我们想得到的目标源声源信号,对于人与人或是人与机器的交流来说,都是相当重要的一件事情。因此盲声源分离在人们的日常生活中是很重要的技术应用。另外盲声源分离技术在人和机器之间的声源通信信道建构也得到广泛应用。盲声源分离(BSS)是一种仅仅使用每个通道观测信号信息的估计源声源信号的方法。我们事先并不需要知道源信号的信息,包括频率特性、源信号空间位置或是源信号是怎样混合的,就可以执行这个估计过程。本文分别研究了声源信号在超定情况下(m?n)卷积混合和欠定情况下(m?n)瞬时混合的盲信号分离问题。具体的研究工作如下:1.首先本文研究了系统是超定的情况下,即从m个卷积混合观测信号中分离出n个源信号。具体步骤为:开始将时域观测信号通过短时傅立叶变换(STFT)转化到时频域。然后利用FastICA分离算法在频域分离观测信号,最后通过频点对换、幅度解混、时频掩蔽和逆短时傅立叶变换一系列操作,我们最终就可以得到估计的源声源信号。2.另外本文研究了系统是欠定的情况下,即从m个瞬时混合观测信号中分离出n个源信号,本文的研究仅考虑了源信号传播过程中幅度衰减和时间延迟的情况,而没有考虑声音混响的情况。分离过程中主要利用了观测声源信号在频域的稀疏性。我们的分离算法是在2路观测信号和3路源信号的实验条件下完成的。具体过程可以分为三个阶段:首先,在频域通过势函数聚类观测信号的角度,将观测信号按它们所属的源进行划分,这时可以估计出衰减矩阵。其次,对于每个划分,通过补偿一个可变时移,我们重新聚类观测信号的角度,直到聚类再次出现,每个可变时移就是时延矩阵的一列,这样我们可以估计出了时延矩阵。最后,通过上面得到的衰减矩阵和时延矩阵,再加上信号频谱系数的幅度是符合拉普拉斯分布这个假设。在混合方程的限定下,我们求得最小的信号幅度和。这实际上是一个二次锥规划的问题。这样我们就估计出了源信号。
[Abstract]:We live in the world of sound. In a noisy environment, it is difficult to get the ideal sound source, and the conversation will become difficult. Therefore, it is very important to get the source of the source of the target from the observed signal with the noise. It is very important for people and people to communicate with the machine. The separation of sound sources is a very important technical application in people's daily life. In addition, the blind source separation technology is also widely used in the sound source communication channel construction between human and machine. Blind source separation (BSS) is a method of estimating source signal only using the information of each channel observation signal. We do not need to know in advance. The information of the source signal, including frequency characteristics, the location of the source signal or how the source signal is mixed, can perform this estimation process. In this paper, the problem of blind signal separation for the instantaneous mixing of the sound source signals in the convolution mixing and the underdetermined case (M? N) under the overdetermined case is respectively studied. The specific research work is as follows: 1. first of all this paper When the system is overdetermined, the N source signal is separated from the M convolution mixed observation signal. The concrete steps are as follows: the time domain observation signal is converted to the time frequency domain through the short Fu Liye transform (STFT). Then the FastICA separation algorithm is used to separate the signal number in the frequency domain, and the frequency shift, the amplitude unmixing and the time frequency mask are finally obtained. In the case of a series of operations by Fu Liye transform and inverse short time, we can finally get the estimated source source signal.2.. In addition, we study that the system is under determined, that is, the N source signals are separated from the M instantaneous mixed observation signal. The study only considers the amplitude attenuation and the time delay in the source signal transmission. There is no consideration of sound reverberation. The separation process mainly uses the sparsity of the observed source signal in the frequency domain. Our separation algorithm is completed under the experimental conditions of the 2 path observation signal and the 3 source signal. The specific process can be divided into three stages: first, the observation in the frequency domain through the potential function clustering observation signal will be observed. The signal is divided according to the source they belong to, then the attenuation matrix can be estimated. Secondly, for each partition, by compensating a variable time shift, we re cluster the angle of the observation signal until the clustering appears again, and each variable time shift is a column of the time delay matrix, so we can estimate the time delay matrix. Finally, pass through The attenuation matrix and the time delay matrix above, and the amplitude of the signal spectrum coefficient are in accordance with the hypothesis of Laplasse distribution. Under the limit of the mixed equation, we get the minimum amplitude of the signal. This is actually a two time cone programming problem. In this way we estimate the source signal.
【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】：TN911.7

【相似文献】