Thank you for your work!
Hallo Mathias,
thanks a lot for making this public! The results are already usable. It's not quite as versatile as Chatterbox's house model is for Englisch, but I'm totally baffled that I can use ALL of my English reference_audio voices templates for German instead! And that they often get way better results from the english input audios than if i use german speech samples. Weird. If i use Stephen Fry (which of course i don't since it would be illegal) it outputs well formed and melodic and sweet flowing german. If i use a german podcaster it becomes HACKED and stilted like Arnold Schwarzenegger. :) Anyway. Thank you for putting in the hard work!
You're welcome. Kartoffel released v0.2 today and I merged my version again with his hard work. I will share.
Hi,
thank you for your work on these models! I’ve tested all of the versions you uploaded and here are my results:
finetuned_by_h2 → sounds completely wrong. Are you sure the correct files were included? I suspect there may be a real issue here.
merged_model → sounds very good.
merged_model_v2 → does not sound good.
For all tests I used German reference voices (in case that has any influence).
From what I see in the files:
t3_cfg.safetensors → is different across all three versions
s3gen.safetensors → is the same between merged_model and merged_model_v2
ve.safetensors → is the same in all three models and sebastian. I assume that is ok :)
That makes me wonder if maybe some files got mixed up or not all were fully uploaded. Could it be that parts of the models were swapped accidentally?
That maybe the reason why merged_model is the only good one!
But the merged_model is very good, thank you again :)
Kind Regards,
Chris