基于自然梯度的语音信号盲源分离优化算法  

Optimization algorithm of blind source separation of speech signal based on natural gradient

作　　者：雷家贵 茹国宝[1] LEI Jiagui;RU Guobao(School of Information and Communication,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学电子信息学院,湖北武汉430072

出　　处：《武汉大学学报（工学版）》2024年第4期528-534,共7页Engineering Journal of Wuhan University

基　　金：国家自然科学基金项目(编号:61671333)。

摘　　要：针对自然梯度算法在语音信号盲源分离中存在的稳定性差、分离性能不佳等问题,利用分离程度作为权重改进Pearson相关系数,将权重改进的相关系数作为控制步长的因子,从而使得步长的变化更加精准,减小随机扰动项对稳态的影响。同时充分利用动量项“奖励一致的梯度,惩罚不一致的梯度”的特性,根据混合信号分离程度的变化方向决定动量项的增减,进一步提高了算法的收敛速度。试验结果表明:与现有的自适应步长自然梯度算法相比,优化算法的信噪比均值更高,方差更低,说明该算法的分离效果更佳,而且减小了对初始值的敏感性,鲁棒性更好;且优化算法的串音误差曲线进入稳态所需的迭代次数更少,说明该算法的收敛速度更快。Aiming at the poor stability and the poor separation performance of the natural gradient algorithm in blind source separation of speech signals,the separation degree is used as the weight to improve Pearson correlation coefficient,and the correlation coefficient whose weight has been improved is used as the factor to control the step size,which makes the step-size change more accurate and reduces the influence of stochastic disturbance term on steady state.At the same time,the momentum term’s characteristics of“rewarding the consistent gradient and punishing the inconsistent gradient”is fully utilized,and the increase and decrease of the momentum term are determined according to the change direction of the mixed signal separation degree,which further improves the convergence speed of the algorithm.The experimental results show that compared with the existing adaptive step size natural gradient algorithms,the algorithm proposed in this paper has a higher mean value of SNR and a lower variance,which shows that the separation effect of the algorithm is better,the sensitivity to the initial value is reduced,and the robustness is better.Moreover,the crosstalk error curve of the algorithm proposed in this paper needs a less number of iterations to enter the steady state,which indicates that the convergence rate of the algorithm is faster.

关 键 词：语音信号分离 自然梯度 自适应步长 相关系数 动量项 

分 类 号：TN911.7[电子电信—通信与信息系统]