Many-to-Many Voice Conversion using Conditional Cycle-Consistent Adversarial Networks
Opened this issue · 0 comments
zbller commented
リンク
https://arxiv.org/ftp/arxiv/papers/2002/2002.06328.pdf
どんなもの？
Added input of spedker's label to CycleGAN
The generator convert voices conditioning on speaker's label
The discriminator distinguish whether the voice is the speaker's one
先行研究と比べてどこがすごい？
Just needed one set of generator and discriminator for n speakers while a basic CycleGAN needs (n-1)・n / 2 sets.
技術と手法のキモはどこ？
Added speaker's label input
どうやって有効だと検証した？
By MOS test, the proposed method got higher score than StarGAN and basic CycleGAN