一种多元非线性回归模型的发音评测技术  

An pronunciation evaluation technology on multiple nonlinear regression model

作　　者：景新幸[1] 熊开芳[1] 杨海燕[1] 

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《大众科技》2012年第12期20-22,59,共4页Popular Science & Technology

基　　金：国家自然科学基金(609661002)

摘　　要：目前,在计算机辅助语言发音评测系统中大部分采用简单的线性回归模型进行机器分与人工评分之间的数据融合,该模型要求机器的评分特征与人的主观评分呈近似的线性关系,而实际呈现非线性。对此,文章引入多元非线性回归模型对分数进行映射融合,使其更符合人的主观评分。同时,在算法优化问题上,用改进步长下的高斯-牛顿算法来求解,大大提高了收敛精度。实验表明,该算法在普通话口语评测中使系统性能得到明显提升。Currently most of computer-assisted language pronunciation evaluation system Integrate data from machine scores and human ratings,using simple linear regression model.The linear regression model requires that the machine was an approximate linear relationship between people＇s subjective ratings and machine scores,while it is actually nonlinear.In response,this paper introduce multiple non-linear regression model for mapping and score fusion,which can make system be more in line with people＇s subjective scoring.At the same time,apply Gauss-Newton algorithm of improved step to solve the algorithm optimization problem and acquire greatly improves in the accuracy of convergence.Experiments show that performance has improved significantly with the algorithm in Mandarin oral evaluation system.

关 键 词：语音评测 多元非线性回归 非线性最小二乘法 高斯-牛顿法 

分 类 号：TN912.34[电子电信—通信与信息系统]