基于改进GMM和韵律联合短时谱的说话人转换  

Voice conversion based on improved GMM and short-time spectrum with prosody

作　　者：张炳[1] 俞一彪[1] 

机构地区：[1]苏州大学电子信息学院

出　　处：《信号处理》2009年第4期548-552,共5页Journal of Signal Processing

摘　　要：提出了一种基于改进GMM模型和韵律联合短时谱的说话人转换方法。通过在训练阶段引入改进的GMM模型,克服传统GMM模型造成的转换语音过平滑现象,并将线谱对频率LSF和基音频率联合起来组成韵律联合短时谱,更准确地刻画说话人的短时频域特征和声腔的共振特性。实验表明,这种方法能够有效地捕捉说话人的个性化特征和韵律特征。另外,在保证变换语音目标倾向性的同时,一定程度上克服了过平滑现象,提高了变换语音的音质。A new voice conversion approach based on improved GMM model and short-time spectrum with prosody was proposed in this paper. And how the over-smoothing phenomenon was alleviated by using improved GMM model was discussed. It was also proposed that LSF and pitch should be put together making up feature vector, which could depict the short-term frequency domain characteristics and the resonance cavity properties of the speaker more accurately, together with the improved GMM model. Experimental resuits show that the algorithm can describe the personality characteristics and prosody features of the speakers more effectively. In addition, it could alleviate the too-smoothing phenomenon and effectively improve the sound quality of transformed speech, while changing the speaker＇s individuality.

关 键 词：说话人转换 改进的GMM 基音频率 韵律 

分 类 号：TN912.33[电子电信—通信与信息系统]