基于一乘准则的LMR在声音转换中的应用  

Application of LMR to Voice Conversion Based on Least Absolute Deviation Criteria

作　　者：邹翔[1] 岳振军[2] 贾永兴[2] 闵刚[1] 

机构地区：[1]解放军理工大学通信工程学院研究生2队,江苏南京210007 [2]解放军理工大学理学院基础电子学系,江苏南京211101

出　　处：《军事通信技术》2008年第1期28-31,共4页Journal of Military Communications Technology

摘　　要：线性多变量回归中聚类数的选择对转换效果有很大影响,通过研究聚类数与类内均方差、类间均方差的关系确定最佳聚类数。针对传统二乘规则的 LMR 的转换语音在听觉上自然度较差的缺陷,提出了基于一乘规则的 LMR,采用遗传算法解决绝对值条件下转换矩阵的求解问题,实验仿真表明,基于一乘规则的 LMR 的转换语音在主观听觉上优于基于二乘规则的 LMR,ABX 测试识别率平均提高了12.5%,MOS 分提高了0.24。The choice of the number of clustering has considerable influence on the effect of conversion:The optimal number of clustering was confirmed by means of the research on the rela- tionship between the number of clustering and the deviation of clustering or between clusterings. According to the limitation of naturalness of transformed speech using traditional LMR based on the Least Squares Deviation Criteria (LSDC),the LMR based on the Least Absolute Deviation Criteria (LADC) was proposed,and the mapping matrix was solved by the Genetic Algorithm (GA).The experiment reveals that the LMR based on LADC has better effect than the LMR based on LSDC,the ratio of identification was promoted by 12.5% in ABX test,and 0.24 was in- creased in MOS.

关 键 词：声音转换 聚类数 最小一乘 遗传算法 

分 类 号：TN91[电子电信—通信与信息系统]