基于空间协方差矩阵的欠定卷积盲源分离

Research on underdetermined convolutive blind source separation based on spatial covariance matrix

利用空间协方差矩阵表示的盲源分离模型与瞬时理想模型的一致性,本文提出了基于空间协方差矩阵的欠定卷积盲源分离方法。本方法用零均值高斯随机变量的协方差矩阵来表示各个源信号经过传输信道后的短时傅里叶变换,采用层次聚类估计出高斯随机变量协方差矩阵的初值,并使用极大期望值算法(EM)求解对数似然函数,最后采用维纳滤波法语音增强技术求解时频域内的源信号。通过仿真实验,验证了算法的有效性。

Based on spatial covariance matrix for utilizing the consistency of the blind source separation model represented by space covariance matrix and instantaneous ideal model, this paper propose a underdetermined convolution blind source separation method. We used the covariance of zero-mean Gaussian random variable to represent the short Fourier transform of each source signal after transmitting in the channel. Adopt the hierarchical clustering to estimate the initial value of the covariance matrix of Gauss random variables. Used the expectation-maximum algorithm to slove the log-likelihood function. At last, using the Weiner filtering speech enhancement technique, figure out the frequency domain source signal. Through the simulation analysisverified the validity of the algorithm.