一种改进的自然梯度语音信号盲分离算法  

Speech signal blind source separation based on an improved natural gradient algorithm

作　　者：岳建杰 赵旦峰[1] 张成[2] 

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]哈尔滨工程大学自动化学院,黑龙江哈尔滨150001

出　　处：《应用科技》2015年第3期30-34,共5页Applied Science and Technology

摘　　要：自然梯度算法有较快的收敛速度、良好的分离性能,在盲信号分离中占有重要地位。基于自然梯度的盲源分离算法一般分为固定步长和变步长的自然梯度算法,固定步长的自然梯度算法存在分离速度与稳定性之间的矛盾,即步长越长时分离速度快,但是稳态误差又得不到保障;步长太小分离速度又达不到要求。为了改善分离速度与稳定性之间的矛盾,提出了一种变步长的方法来,并用其改进了固定步长的标准自然梯度算法,成功地用于混合语音信号的分离,该方法取得比标准自然梯度算法更好的分离效果,具有更快的收敛速度。A natural gradient algorithm has a fast convergence rate and excellent separation property, and thus plays an important role in blind signal separation. The blind source separation algorithms based on natural gradient are commonly classified into fix-step-size and variable step-size algorithms. The fix-step-size natural gradient algorithm has the inherent contradiction between the speed of separation and the error in steady state. Namely the bigger the step size is, the faster the separation speed is, but the error in steady state cannot measure up. Then a small step size would lead to slow separation speed. In order to improve the contradiction, this paper proposes a new method based on variable step-size natural gradient, which is used to improve the fixed-step-size standard natural gradient algorithm and it has been successfully applied to separate mixed speech signals. The new algorithm has gained bet-ter separation performance and faster convergence speed in inter symbol interference than the standard one.

关 键 词：语音信号处理 自然梯度算法 收敛速度 盲信号分离 

分 类 号：TN971.1[电子电信—信号与信息处理]