摘要
为了分离具有时序结构的信号,将线性预测均方误差作为代价函数,使分离出信号的可预测性最大,这样就可以分离出源信号。这种最小均方误差型算法,其在线形式采用瞬时预测误差代替预测误差的期望值,导致收敛速度较慢。为了提高这类算法的收敛速度,本文将线性预测误差的加权平均作为代价函数,提出了递归最小二乘型线性预测盲源分离算法。计算机仿真和实际语音分离试验均表明:提出的算法与最小均方误差型线性预测盲源分离算法相比具有更快的收敛速度,且增加的计算量不大。
Mean square error of the linear predictor is used as the cost function to maximize the predictability of the extracted signal so that a source signal with temporal structure can be extracted. But this algorithm （called the LMS-LP-BSS） is a least mean square type, instantaneous prediction error is used instead of expectation of the prediction error in the online version of the algorithm, resulting in low convergence rate. To improve the convergent speed of the algorithm. The weighted mean error of the linear prediction is used as the cost function and recursive least square algorithm and the linear prediction for blind source separation （RLS-LP- BSS） is proposed. Simulations and experiments on extraction of real speech voices show that the proposed algorithm has high convergence rate than LMS-LP-BSS with a little more computation.
出处
《数据采集与处理》
CSCD
北大核心
2008年第1期60-64,共5页
Journal of Data Acquisition and Processing
关键词
递归最小二乘
线性预测
盲源分离
recursive least squares
linear prediction
blind source separation
