改进的GMM模型语声转换系统  

A Modified Method for Voice Conversion Based on GMM

作　　者：申毅[1] 简志华[1] 杨震[1] 

机构地区：[1]南京邮电大学信号处理与传输研究院,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2007年第5期11-15,21,共6页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：江苏省"青蓝工程"基金(QL003YZ)资助项目

摘　　要：传统的基于GMM模型线性语音转换系统在语音转换阶段,由于转换函数的概率加权组合使合成语音共振峰带宽变宽,谱包络过于平滑。文中提出依据后验概率大小和前后语音的相关性,选择部分转换分量函数进行语音转换。实验表明不仅简化了语音转换,而且经过转换的语音质量也有一定的提高,对语音的实时转换有重要的意义。　For the traditional GMM-based linear voice conversion system,due to the probability weighted combination of conversion function,the resonant peak width of composite voice is broaden and the spectral envolop is flat.The autors propose to convert voice using partral conversion component function according to posterior probability and correlation of adjacent voice signals.The experiments have prove that the voice conversion is simplified and the converted voice quality is improved.It is important for real-time conversion of voice.

关 键 词：语音处理 语声转换 韵律转换 高斯混合模型 

分 类 号：TN912.3[电子电信—通信与信息系统]