针对语音信号的卷积混迭模型，通过对接收信号进行滑窗处理并重新排列，使得重构后的接收信号可表示为块独立的源信号与扩展的混迭矩阵的瞬时混迭模型。利用不同语音信号之间的近似独立和短时平稳特性，分析接收信号二阶相关矩阵的非正交联合块对角化结构。改进基于Givens和Hyperbolic旋转的瞬时混叠盲分离算法，提出Givens-Hyperbolic双旋转的联合块对角化算法(GH-JBD)。所提出的GH-JBD算法可以直接在时域估计传输信道的块本质相等矩阵，实现非正交联合块对角化，进而完成卷积混叠信号盲分离。该方法不需要进行预白化处理，避免了由于白化不彻底而引入的额外误差。仿真实验验证了该算法的有效性并就参数变化对分离信号的影响进行了分析。%The convolutive mixture model of speech signal could be expressed as an instantaneous mixture of block-independent source signal and extended mixture matrix by reconstructing the received signal after sliding-window segmentation. And based on the mutual-independence property and the short-time stationary of the speech signals, the second-order correlation matrix of the received sig-nal has a structure of non-orthogonal joint block-diagonalization. Then a new Givens-Hyperbolic double-rotation based joint block-diagonalization algorithm GH-JBD is proposed by improving the existing algorithm of instantaneous mixture blind separation based on Givens-and-Hyperbolic rotations. The GH-JBD algorithm estimates the block essentially-equal matrix of the mixture matrix directly in time domain to realize non-orthogonal joint block diagonalization for convolutive blind source separation. Since pre-whitening of the received signal is not needed, there is no residual error induced in GH-JBD processing. Simulations prove the validity of the proposed algorithm in various scenarios.
展开▼
