部分W-分离正交语音信号的盲分离方法

Blind Separation for Partial W-Disjoint Orthogonal Speech Signals

在W-分离正交性假设的语音盲分离方法中,由于没有考虑多个源信号同时存在的情况,导致分离信号中不可避免地存在音乐噪声。针对这种部分W-分离正交情况,提出了基于信道估计的语音盲分离方法。该方法先检测只有一个源信号存在的时频点并进行归一化处理,使得处理后的结果与频率无关,克服了W-分离正交性假设的不足以及频率置换问题,通过K-means聚类估计出信道,再结合信号子空间方法重构源信号。仿真结果表明,提出的方法可以有效减少分离语音中的音乐噪声,与典型的时频二元掩蔽方法相比,其平均信号失真比提高3.02dB,同时平均信干比提高4.61dB。

In blind speech separation methods based on the assumption of W-disjoint orthogonality （W-DO）, musical noise is inevitable in separated signals because the assumption does not include the case of existing multiple source signals in the time-frequency domain. A blind speech separation method based on channel estimation is proposed for partial approximate W-disjoint orthogonality. The time-frequency cells with only one source are detected and normalized to be independent of frequency, which overcomes not only the shortcoming of W-DO property but also the frequency permutation problem, and then the channel estimation is obtained by K-means clustering. Finally, signal subspace method is exploited to reconstruct sources. Simulation results demonstrate that the novel method can effectively reduce the musical noise in the separated speech signals, and it outperforms the typical time frequency binary masking method, the averaged signal to distortion ratio （SDR） is improved by 3.02 dB and the averaged signal to interference ratio （S/R） is improved by 4.61 dB.