StarGANv2-VC: A Diverse, Unsupervised, Non-parallel Framework for Natural-Sounding Voice Conversion

We present an unsupervised non-parallel many-to-many voice conversion (VC)\nmethod using a generative adversarial network (GAN) called StarGAN v2. Using a\ncombination of adversarial source classifier loss and perceptual loss, our\nmodel significantly outperforms previous VC models. Although our model is\ntrained only with 20 English speakers, it generalizes to a variety of voice\nconversion tasks, such as any-to-many, cross-lingual, and singing conversion.\nUsing a style encoder, our framework can also convert plain reading speech into\nstylistic speech, such as emotional and falsetto speech. Subjective and\nobjective evaluation experiments on a non-parallel many-to-many voice\nconversion task revealed that our model produces natural sounding voices, close\nto the sound quality of state-of-the-art text-to-speech (TTS) based voice\nconversion methods without the need for text labels. Moreover, our model is\ncompletely convolutional and with a faster-than-real-time vocoder such as\nParallel WaveGAN can perform real-time voice conversion.\n