基于多领域条件生成的语音情感转换

Emotional Voice Conversion Based on Multiple Domain Conditional Generation

语音情感转换是在不改变话者声纹、语义的情况下,将一种情感语音转换成另一种情感语音的技术,本质是实现语音的风格迁移。主流的风格迁移技术有对抗生成技术(如CycleGAN,StarGAN)和实例规一化技术(如IN,CIN)。CIN相对于IN添加了均值方差选择性模块,具有更强的风格迁移能力。提出了将StarGAN和CIN结合的语音情感转换模型CIN-StarGAN,将CIN模块嵌入到StarGAN生成器。在ESD数据集上的实验结果表明,CINStarGAN比基于CycleGAN的情感转换模型收敛速度快28%,具有较好的风格转换能力。在多领域情感转换方法上具有潜在研究价值。

Emotional voice conversion was a technology that converted the emotion of a speech into another without changing the speaker′s timbre and semantics.Its essence was to transfer style of speech.The mainstream style transfer technologies included generative adversarial network(such as CycleGAN,Star-GAN)and instance normalization technology(such as IN,CIN).Compared with IN,CIN added a mean variance selective module,which had stronger style transfer ability.StarGAN and CIN were combined,and proposed a new speech emotion conversion model,CIN-StarGAN.The model embeded the CIN module into the StarGAN generator.The experimental results on ESD data sets showed that CIN-StarGAN converged 28%faster than CycleGAN based emotion conversion model,and had better style transfer ability.It had potential research value in multi domain emotion transfer methods.