Many-to-Many Voice Conversion using Conditional Cycle-Consistent Adversarial Networks

Voice conversion (VC) refers to transforming the speaker characteristics of\nan utterance without altering its linguistic contents. Many works on voice\nconversion require to have parallel training data that is highly expensive to\nacquire. Recently, the cycle-consistent adversarial network (CycleGAN), which\ndoes not require parallel training data, has been applied to voice conversion,\nshowing the state-of-the-art performance. The CycleGAN based voice conversion,\nhowever, can be used only for a pair of speakers, i.e., one-to-one voice\nconversion between two speakers. In this paper, we extend the CycleGAN by\nconditioning the network on speakers. As a result, the proposed method can\nperform many-to-many voice conversion among multiple speakers using a single\ngenerative adversarial network (GAN). Compared to building multiple CycleGANs\nfor each pair of speakers, the proposed method reduces the computational and\nspatial cost significantly without compromising the sound quality of the\nconverted voice. Experimental results using the VCC2018 corpus confirm the\nefficiency of the proposed method.\n