Voice Conversion Using Conditional CycleGAN

Voice conversion (VC) modifies characteristics of speech, such as gender and speaker identities. The VC can be applied to various tasks including speaking assistance and speaker anonymization. Generally, such VC techniques require parallel speech data for training, which is very expensive. Recently, voice conversion has been accomplished using CycleGAN, which does not require parallel speech data. In this paper, we further extend the idea of using CycleGAN to convert multiple speakers' voices by conditioning the CycleGAN using speaker identity information.