基于分类线性加权的源-目标话者声音转换算法的研究  

Voice conversion from source speaker to target speaker based on classified linearly weighted transformation

作　　者：张剑[1] 戴蓓蒨[1] 孙俊[1] 陆伟[1] 李辉[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,安徽合肥230026

出　　处：《电路与系统学报》2008年第3期106-110,105,共6页Journal of Circuits and Systems

摘　　要：源-目标话者的声音转换是一种变换说话人声音特性的技术,它将源说话人的声音转换成另一个指定的目标说话人的声音。对源话者声道谱特性的修改是声音转换的关键之一。为了克服一般分类线性转换算法中分类不准确所带来的误差,本文引入了分类线性加权转换的策略,根据不同子类的转换函数对谱特性的贡献,赋予不同的加权系数,给出了一种基于GMM后验概率加权的线性转换算法。在微软汉语普通话语音数据库上做的四组对比实验表明,该算法在谱转换性能上均有不同程度的提高。voice conversion technique aims to modify the source speaker＇s speech to make it sound like a designated target speaker＇s speech, of which the spectral envelope mapping algorithm is the key part. A classified linearly transformation is introduced to reduce transformation error caused by inaccurate classification. Different weighted values are added based on the contribution of each class to the whole spectral envelope, and a weighted linearly transformation based on the GMM posterior probability is presented. Experimental results show the proposed algorithm can improve the performance of converted spectral envelope.

关 键 词：声音转换 源-目标话者 声道谱转换 高斯混合模型 分类线性转换 分类线性加权转换 

分 类 号：TN912.34[电子电信—通信与信息系统]