基于生成对抗网络的声音迁移技术应用  

Application of Sound Migration Technology Based on Generative Adversarial Networks

作　　者：罗钇凯 周骁 俞可扬 沈启承 户希娅 吴子朝[1] LUO Yikai;ZHOU Xiao;YU Keyang;SHEN Qicheng;HU Xiya;WU Zichao(Hangzhou Dianzi University,Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学,浙江310018

出　　处：《电子技术（上海）》2023年第3期94-95,共2页Electronic Technology

摘　　要：阐述一种声音风格迁移的技术系统,该系统具有较强的灵活性,进行风格迁移的对象不仅仅包括人声,甚至可以是音乐风格。此处声音风格迁移指的是对于指定人向指定人的语音风格转移技术。此外该方法不依赖于并行数据,并且可以对任意长度的样本进行转换。该技术采用生成器-判别器的GAN架构并加以对抗性约束,由此生成高度逼真的样本。This paper presents a technical system of sound style transfer,which has strong flexibility.The object of style transfer is not only human voice,but also music style.Here voice style transfer refers to the technique of transferring voice styles from person to person.In addition,the method does not depend on parallel data and can transform samples of any length.This technique uses a generator-discriminator GAN architecture with adversarial constraints to generate highly realistic samples.

关 键 词：语音迁移 GAN架构 对抗性约束 

分 类 号：TN912.3[电子电信—通信与信息系统]