基于平均音素模型的音色转换研究  

Research on Voice Conversion Based on Average Phoneme Model

作　　者：赵薇[1] 唐堂 ZHAO Wei;TANG Tang(School of information and Communication Engineering,Communication University of China,Beijing 100024,China)

机构地区：[1]中国传媒大学信息与通信工程学院

出　　处：《中国传媒大学学报（自然科学版）》2020年第1期1-6,共6页Journal of Communication University of China：Science and Technology

基　　金：国家自然科学基金(61901421);中央高校基本科研业务费专项资金(CUC19ZD003)

摘　　要：音色转换技术能够在保留原有语句信息的基础上,使原说话人的声音特征向着目标用户的声音转变,从而达到用目标用户声音替换说话人声音的目的。在训练目标人音色时,传统方法需要大量的语料库进行训练。但是制作语料库花费很多的时间及人力,无法满足音色模板快速生成的需求,在实现个性化音色灵活性方面受到限制,很难扩展或显著改进。本文利用praat软件提取语音音素,通过GMM-UBM系统训练平均音素模型,利用较少的语音数据训练,从而实现在短时间小样本情况下个性化音色模型的建立,完成音色转换。主观实验表明,该方法达到了很好的音色转换效果。Voice conversion technology can change the original speaker's voice characteristics to the target user's voice on the basis of retaining the original sentence information,so as to achieve the purpose of replacing the speaker's voice with the target user's voice.When training the target person's timbre,the traditional method needs a large number of corpus for training.However,the production of corpus takes a lot of time and manpower,which cannot meet the needs of rapid generation of voice template.It is limited in the realization of personalized voice flexibility,and it is difficult to expand or significantly improve.In this paper,Praat software is used to extract speech phoneme,GMM-UBM system is used to train average phoneme model,and less speech data is used to train,so as to realize the establishment of personalized voice model in a short time and small sample,and complete voice color conversion.Subjective experiments show that this method achieves a good effect of timbre conversion.

关 键 词：音色转换 praat软件 GMM-UBM 平均音素模型 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]