论文部分内容阅读
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 selection of the number of clusters in linear multivariate regression has a great influence on the transformation effect. The optimal cluster number is determined by studying the relationship between the number of clusters and the mean variance within the cluster and the mean variance between clusters. In order to solve the problem of solving the conversion matrix with absolute value, a genetic algorithm based on one-rule LMR is proposed to solve the defect that the converted speech of traditional two-rule LMR is less aurally natural. Experimental results show that the LMR based on one- The rule-based LMR conversion speech is subjectively superior to the LMR based rule of squares, with an ABX test recognition rate of 12.5% on average and a MOS score of 0.24.