ai model quality problem

[split nova]

Hello, I would like to ask for help
I made an AI model but there are errors (the sound cracks/drops out at high notes and there is a lot of noise during the breaks in the music) and out of curiosity I tried what happens if I convert it to an image and compared it with downloaded models
and it appeared at the beginning and at the end in my case there is something that sounds like noise, while at the beginning there is complete silence in the original file
the downloaded ones don't have it
and I would like to know how to make it disappear

or anything that could be better

(if everything is true, the first picture is the original 20hz-20khz sound, the 2 after it are mine and the last 2 are from another model)

[rapid mirage]

is that 500+ epochs?

this test can evaluate the dataset

we use this file

so you're training a 48k with 32k dataset

[split nova]

This is the e500 model I made
I did everything in the training as written on the applio website

Should I make it with the rest?

[rapid mirage]

you're training 32k dataset in 48k, which lets the model learn useless noise

the dataset is lacking the singing frequencies (>15s curve is missing)

500e for a finetune is extemely excessive

[split nova]

so should the source sound that it learns from be exported like this?
the one I made doesn't usually sing, I can only get the speech from youtube
and he made it up to 500e, I don't know how to fine-tune it

[split nova]

of course in mono

[rapid mirage]

there's not much reason to use higher SR model training when your dataset does not have such range

[split nova]

In this case, how big should it be?

How should I do it? Could you explain more?

Should both be 32k or should they just be the same?

[rapid mirage]

if your dataset range does not go over 16k on spectrogram, then you should train the model at 32k (16k x 2)

[split nova]

so far I understand
and can the file remain 48k or should I export it in 32k too?

and how much should I use instead of 500e?
currently I have it on default settings 12 10 500

[rapid mirage]

first you may want to update applio to latest version 3.3.0

you dont need to resample your dataset, applio does it for you

[split nova]

and one more question
this person also has another voice because after a windows reinstallation the voice was set differently
if I want to can I add that to the data set or should I make it as a separate model?

[rapid mirage]

idealy for the dataset you want the voice recorded in the same conditions/volume/etc

[split nova]

Okay then that will be separate

and how should I set it here?
I've left it on default so far

[rapid mirage]

batch size depends on the dataset size (and limited by VRAM from the top)

max epoch for finetuning should not really go over 300... 500

the model done when tensorboard loss charts converge.. then you test the saved models and pick the best

[split nova]

Thank you very much
I'll try it again

[tame axle]

I'd suggest to provide the dataset in the sample rate same as the target (i.e. 32k for your case)

also the pretrain should match the target sample rate (the default one will adapt to it)

[split nova]

now it's like this