search

GANtastic3 / MaskCycleGAN-VC

Implementation of Kaneko et al.'s MaskCycleGAN-VC model for non-parallel voice conversion.
MIT License
112 stars 31 forks source link

Different source speakers #14

Open EmreOzkose opened 3 years ago

EmreOzkose commented 3 years ago

Hi, thanks for sharing such a good work!

Have you ever tried training with different source speakers?

sauravpd29 commented 3 years ago

@EmreOzkose how much time did it take for training? You used GPU?

EmreOzkose commented 3 years ago

I haven't tried training with different source speakers yet.

EmreOzkose commented 3 years ago

I am doing some experiments with different dataset. Results are very good if I use single speaker to target. However when I use multiple source speaker to one target, it is not good as previous one. I just wanna report. Tensorboard output is here:

Screenshot from 2021-07-01 18-22-35

Green one is continuation of blue one. Blue -> multi-speaker Orange -> single-speaker

Multi-speaker discriminator cannot decrease after a point. (~1M)