breaking the training barrier #6
Replies: 4 comments 4 replies
-
|
Yeah, I know—really awesome, right? I stumbled upon this a few days ago because I thought it would be easy and interesting enough to do my first model fine-tuning on something I really wanted. What a blast! |
Beta Was this translation helpful? Give feedback.
-
|
@dida-80b Btw. My Stage 2 Epoch 1 finally finished and it sounds great. Not finished but still great. Have re verified that the adversarial training in the later epochs of Stage 2 are actually working? I have not made it there yet. |
Beta Was this translation helpful? Give feedback.
-
|
Yes, I can confirm... I made it to Stage 2 Epoch 4, with the adversarial training kicking in from Epoch 3 (joint_epoch=3). The losses all looked healthy throughout: Disc Loss ~3.97, Gen Loss ~3.7, WavLM Loss ~1.5. No instability, no collapse. And subjectively the audio quality did improve noticeably in the later epochs — Eva-K's voice character comes through more clearly and the speech feels more natural compared to the earlier epochs. Still room to grow, but the adversarial training is definitely doing its job. |
Beta Was this translation helpful? Give feedback.
-
|
@semidark i suggest you make a reddit post. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
we have essentially established the first fully functional, publicly accessible training workflow for kokoro. this isnt just about german. these fixes provide a blueprint for anyone looking to train kokoro for any language.
Beta Was this translation helpful? Give feedback.
All reactions