论文部分内容阅读
针对语音信号的欠定卷积混合模型,利用独立语音在时频域上的近似W-分离正交性(W-DO),提出了一种基于非线性时频掩蔽的盲分离方法。首先对多传声器观测信号在时频域上进行规范化处理,使混合信号在每个时频槽的表示与频率无关,然后采用动态聚类算法获取时频槽对应的活跃源信息,选择关于簇中心偏角的非线性函数进行时频掩蔽,从而实现语音信号的盲分离。该方法解决了经典频域盲分离算法中的频率置换问题,能有效抑制分离矩阵的空间方向扩散。仿真实验表明,与BLUES方法相比具有更优的分离性能,信噪比增益平均增加1．58 dB。
Aiming at the underdetermined convolutional hybrid model of speech signals, a blind separation method based on nonlinear time-frequency masking is proposed by using the approximate W-DO (Orthogonality) of independent speech in the time-frequency domain. Firstly, the multimonitor observations are normalized in the time-frequency domain so that the mixed signal is independent of frequency in each time-frequency slot. Then, dynamic clustering algorithm is used to obtain the active source information corresponding to the time-frequency slot. Angular non-linear function of time-frequency masking, in order to achieve blind separation of speech signals. This method solves the problem of frequency substitution in classical frequency domain blind separation algorithm and can effectively suppress the spatial direction diffusion of separation matrix. The simulation results show that compared with the BLUES method, the signal to noise ratio gain increases by 1.58 dB on average.