一个用于卷积混合模型的时频分析盲分离算法

Blind Source Separation Algorithm for Convolutive Mixtures Based on Time-Frequency Analysis

提出一种基于时频分析的卷积混合盲分离算法。由于信号源与各传感器的距离不同,在传播的过程中会产生不同的幅度衰减和时间延迟。该算法用短时傅里叶变换对语音信号进行时频分析,将其中一个传感器信号作为参考信号,构造了源信号的幅度衰减向量和时间延迟向量。根据语音信号的时频域稀疏性,以这两个向量为特征,在时频域上对传感器信号进行聚类,再通过估计参考信号混合系数来获得源信号时频域表示,进一步得到源信号。该方法可以用于源信号数目大于传感器信号数目的情况。仿真实验证明,算法可以完成欠定情况下卷积混合信号的盲分离,分离结果令人满意。

A speech blind separation algorithm for convolutive mixtures based on time-frequency analysis is proposed. Because the distances between sources and sensors are different, am- plitude attenuations and time delays can be produced with the source signals transmit. The algorithm uses the short-time Fourier transform to process the sensors signals, and constructs the amplitude attenuation vectors and time delay vectors. Using the speech signal sparsity, mixtures in time-frequency domain can be clustered into several classes according to amplitude attenuation vectors and time delay vectors. The representations of source signals in time-fre- quency domain are obtained by estimating the mixed coefficients. The algorithm is used for the case of more sources than sensors. The simulation shows that the algorithm can separate con- volutive mixtures in undetermined case, and results demonstrate encouraging separation perfor- mance of the signals.