Korean Voice Conversion using Cycle-GAN.
It is tensorflow 2.0 implementation of the paper : "CycleGAN-VC Parallel-Data-Free Voice Conversion Using Cycle-Consistent Adversarial Networks" (I changed some hyper-parameters to apply to Korean-dataset)
You can listen converted voices in Sample directory
- librosa
- pyworld
- tensorflow 2.2
- tensorflow addons
Download Korean dataset here.(https://ithub.korean.go.kr/user/total/referenceManager.do)
I used Man's voice and Woman's voice in 20s.
Locate train dataset and validation dataset to designated folders.(train dataset : train_A, train_B / validation dataset : validation_A, validation_B). Before starting training, preprocessing is carried out. (It may take quite a long time)
Trainging takes at least 100 epochs to achieve good quality results.
!python main.py
Locate test dataset to designated folders.(test_A, test_B)
!python test.py
- https://arxiv.org/abs/1711.11293
- https://github.com/pritishyuvraj/Voice-Conversion-GAN : pytorch implemantation. I referenced the preprocessing code.