Open Ph0rk0z opened 1 year ago
They might be better, i'll implement it. Not immediately though, i'm currently making an improved install system.
No worries. For install I actually reused my nvida environment and added the handful of extra packages so I always bypass the venv entirely.
I'd like to request adding the option of using the "mango-crepe" algorithm too for training and generating
While adding v2 48k i realized v1 48k doesn't work, i have some fixing to do
Yea.. you're right. I had tried V1 48k and assumed something was wrong with my system instead.
Yeah, once i fix that, v2 48k should also work
I will try to re-run some datasets and see if the quality is higher. If only any demucs supported 48k. With already clean samples or self-recorded audio it will probably be a nice improvement.
48k failed for me on training. I got the d/s converted. The 2nd wav folder doesn't generate files. The one where the # of items would be double. All the other ones happen but training goes down on tensor shape/size.
48k failed for me on training. I got the d/s converted. The 2nd wav folder doesn't generate files. The one where the # of items would be double. All the other ones happen but training goes down on tensor shape/size.
That message above was one message not 2, it said once i fix that, v2 48k should also work
, but i haven't worked on fixing it yet.
Fair, I'm just saying how far I got testing it.
48k failed for me on training. I got the d/s converted. The 2nd wav folder doesn't generate files. The one where the # of items would be double. All the other ones happen but training goes down on tensor shape/size.
That message above was one message not 2, it said
once i fix that, v2 48k should also work
, but i haven't worked on fixing it yet.
Hi Mylo!, is v2 48k working currently or shall we wait?
48k failed for me on training. I got the d/s converted. The 2nd wav folder doesn't generate files. The one where the # of items would be double. All the other ones happen but training goes down on tensor shape/size.
That message above was one message not 2, it said
once i fix that, v2 48k should also work
, but i haven't worked on fixing it yet.Hi Mylo!, is v2 48k working currently or shall we wait?
Not yet, i haven't really been working on it, 48k works in inference by the way, just not during training.
There are now 48k v2 models and a fork with a hybrid training feature. Are they any better?
https://github.com/Mangio621/Mangio-RVC-Fork
Files from YT are generally 48k. The only issue is the demucs models mainly do 44.1, although demucs can supposedly support 24bit. https://github.com/facebookresearch/demucs/issues/288
Am unsure how upsample vs downsample works out for final output.