Never seen such behavior before, but the harmonics seems 'mirrored' everywhere.
This is most severe in singing inferences whereas less in speech, not sure what caused this so yea...
Was using RMVPE & og pretrain.
#'Mirrored' harmonics from trained model
glossy cape
low cedar
Never seen such behavior before, but the harmonics seems 'mirrored' everywhere.
...
I thought it may happen more on 32k models than 40k as ever discussed in https://discord.com/channels/1159260121998827560/1265083834039533588
my possible guess might be overly denoising to the point of sounding a bit muffled
low cedar
u can also check some discussions in #🔊│ai-development message
glossy cape
I thought it may happen more on 32k models than 40k as ever discussed in https:/...
Thanks for the advise! I will try to re-denoise/remove those that sound muffled.
Also, is it possible that the dataset itself has such issue? I've found some portions that looks sus to me (Not that outstanding when listening to it):
bright path
Thanks for the advise! I will try to re-denoise/remove those that sound muffled....
dont train from scratch, use a default pretrain
glossy cape
dont train from scratch, use a default pretrain
I was using default pretrain
bright path
hold on
how many epochs was that your original picture and do you have the audio you used for that?
I've never seen such clear lines on the spectrogram, and no such crazy mirroring
48k vs 32k models based on the same set
48k is 120 epochs, 32k is 250 epochs
frail jewel
48k vs 32k models based on the same set
Which would be better, 32k, 48k or 40k using the original train?