Girl | RMVPE | 90 Epochs

[lunar reef]

Woman

[surreal condor]

damn 90 epochs and this good?

[lapis zealot]

More Epochs doesnt mean the model will be better. I was also suprised that this model sounded good at that checkpoint.

[surreal condor]

doesn't more epochs mean the amount of rounds the model went thru??

[lapis zealot]

1 epoch is the model going through the whole dataset once. So for this model it went through the entire dataset 90 times.

[surreal condor]

so why more doesn't mean better? I mean yeah over training and stuff

[lapis zealot]

Overtraining. As far as i know thats the only reason.

[teal slate]

model starts to forget the pretrain previous knowledge and proceeds to generate trash

[surreal condor]

yeah but most models are not that good at 90 epochs hence my amusement

[surreal condor]

I never knew this

[lapis zealot]

That would be a dataset problem, no?

Longer dataset needs a less epochs?

[surreal condor]

I mean most of my datasets are 42 min

and I run them 350 epochs

[lapis zealot]

Do you check for overtraining at all?

[surreal condor]

using tensor board

[lapis zealot]

Why is this not true Lyery?

[teal slate]

ideally you want a sweetspot where the model has learned a considerable portion of your dataset, but at the same time not forgetting the pretrain

[surreal condor]

I use the pretrains that came with rvc.
-# I don't know which to choose for what

[teal slate]

you cannot predict how many epochs it's going to take for a model to extract something useful, thats for every dataset regardless of size

[lapis zealot]

If only the TensorBoard had any validation metrics i think it would be pretty useful, at this current time in my opinion it doesnt tell you much.

[lapis zealot]

I see, thanks.

[teal slate]

the original rvc-dev recommended 20-40 epochs

up to a max of 200 if the dataset has a lot of features

back then the ui limited the max epoch amount to 100

[surreal condor]

I heard that it there are differnt pretrains for differnt scenarios

like singing

e-girls

asmr

[teal slate]

true, but they're all outdated
i'd recommend either only the og pretrain or klm

i do have a spin pretrain only for realtime models

which uses the same dataset as og

[lapis zealot]

Spin?

[teal slate]

a new embedder, handles single words better than cvec and handles context differently (in a better way)

[surreal condor]

what is the best way to get a model better at laughting and breathing
I mean my dataset had a LOT of breathing and coughting and lauhging and still there are a lot of artifacts
I save every 50 epochs and I try them all and all of them sound weird

[teal slate]

cvec relies too much on context, it really needs to see the whole file in order to produce correct speech

[lapis zealot]

If its new and better why is it not standard yet?

[teal slate]

which is ok for local conversions but on realtime it struggles so hard even when extra is set to max

[teal slate]

no many pretrains done for it

[lapis zealot]

Where is the one you made?

[teal slate]

some finetune i did yesterday with it

works fine

it can't sing tho, its only meant for speech

[lapis zealot]

Thanks!

[lapis zealot]

That is fine.

[surreal condor]

wait the model was trained with english language and was still able to create this?

[teal slate]

yes

[teal slate]

spin models only work with vonovox!

doesn't work in w-okada or mainline realtime

[surreal condor]

that would be huge if it works ir time for my german language

[lapis zealot]

Ok, i use vonovox anyway so thats fine.

[teal slate]

[teal slate]

spin was also trained using only english, but i've heard it has improved pronunciation in other languages compared to cvec

only way to know is to try it

[surreal condor]

but does it work live on voice changers like okada?

[teal slate]

only vonovox supports it
but it's possible to use in w-okada by doing this https://github.com/deiteris/voice-changer/pull/213

[surreal condor]

hmmm lemme try

[teal slate]

i have no idea which spin embedder is that tho, my spin pretrain only works with layer 7_12* (same used in vonovox and applio)

[lapis zealot]

Unrelated question but, do you know of any script, fork, or something i can do to get some sort of validation in rvc?

[teal slate]

codename's fork has pesq validation alongside si_sdr

it only validates the same voice tho

[lapis zealot]

Do you have a link to that?

[lapis zealot]

Better than nothing.

[teal slate]

bruh

sdf forgot to copy and paste the link

this

[lapis zealot]

Thank you.

[teal slate]

have you heard of utmosv2? validates how natural/human a spectogram looks

[lunar reef]

The only "good" pretrain for now js the original, Titan or Ov2 for example provides robotic results for kinda no reason

[lapis zealot]

I have not, i am still new to ai audio.

[teal slate]

https://github.com/sarulab-speech/UTMOSv2

gives a MOS score to your generated ai audio files

in a scale of 1 to 5

serves as post validation

code tried to implement it in his fork but failed due to utmosv2 not working properly with multiprocessing

[lapis zealot]

Thats unfortunate.

Seems useful.

[teal slate]

place this in utmosv2 > utils

fixes the broken auto downloader

my infer settings: python inference.py --input_dir "input folder" --out_path "C:\Users\xxx\Desktop\results\results.csv" -n 0 --fold -1 --num_repetitions 20

[lapis zealot]

What does --fold -1 mean? In looking in the code and there doesnt seem to be a comment describing it.

Unless im missing it.

[teal slate]

uses all pretrained models during inference

for more accurate results

[lapis zealot]

Ah ok thanks.

[boreal fable]

what about original wokada? is he really doing only UI changes lol

[teal slate]

did he added spin support?

[boreal fable]

I haven't checked, it was actually question since with "wokada" you seem to only mean wokada deiteris fork

i actually got 2.1.4 alpha for a test for noobies