[Pretrains Test] Ariana Grande (RMVPE) (RVC v2, 32k, 200 epochs, 8400 steps)

[visual beacon]

I made these models to test the newest pretrains
Link: https://huggingface.co/shiromiya/public-models/tree/main/ariana-grande-pretrains-test
Dataset time: 54:00
Trained on Applio v3.2.5
List of Pretrains:

• Original
• Titan Medium
• Rigel FT
• Ov2
• Nanashi v2 FT
• DMR v1
• KLM 4.1
• Singer pretrain

If you want me to do it on another pretrain, leave a comment with the name of the desired pretrain

[maiden pendant]

Rate the model

To rate this model, either use /rate or click the buttons below.

[visual beacon]

Samples:

I'm training with DMR now

[junior walrus]

without looking the spectograms and only using my ears, for me the original sounds better

i order them from best to bad based in my opinion and ears xd:

1. original pretrain
2. titan
3. nanashi
4. ov2
5. rigel

[sour ore]

Shiry, you used studio stems or vocal takes from protools sessions?

[sour ore]

I agree, OG sounds the best and the worst ones are nanashi and rigel.

(No offense Shirou)

[visual beacon]

Studio steams

[visual beacon]

How is Ov2 worse than Nanashi? lol

[sour ore]

Idk if it's my personal liking but i prefer to use raw vocal takes.

[junior walrus]

i ordered them from best to worst
nanashi is above ov2 lols

for me the worst are ov2 and rigel

[visual beacon]

Well, just look at the warning on Nanashi's post and you'll get an idea of my opinion of Nanashi

[junior walrus]

but i would say ov2 and nanashi are extremely similar, maybe in the same tier
ov2 has better breaths tho

(again, only hearing them and not using the spectogram to compare them)

[visual beacon]

Ah I see

[visual beacon]

The problem with them was that they had some background noises, so I decided to leave them out lol

[sour ore]

BS Roformer and RX10: Allow us to introduce ourselves.

https://tenor.com/view/allow-us-to-introduce-ourselves-veggietales-veggie-larry-the-cucumber-bob-the-tomato-gif-21074601

[visual beacon]

Laziness lol

[vocal mortar]

Ayo? @visual beacon level 9 !!!

[rigid forge]

Bacht size used?

[visual beacon]

16

[rigid forge]

wow

[rigid forge]

because i would like to rifarte vcon original and 250 epochs serj tankian of system of a down

[rigid forge]

Is it okay at 16 for him?

[rigid forge]

in the sense according to you

[visual beacon]

I think so, I've done smaller models with this BS and it worked well

[visual beacon]

DMR v1 added
Sample:

[terse crater]

Why 200 epochs? Is it just an arbitrary number?

[visual beacon]

I chose this value because it's 54 minutes of dataset, and I didn't want to spend a week doing it lol

[terse crater]

I see

[terse crater]

Heres a some of info based on the spectrogram. Best to worst:

• DC offset:
  • DMR
  • OV2
  • SingerPretrain
  • TITAN, OG
  • Rigel
  • Nanashi
  • KLM 4.1
• Harmonic Distortion:
  • Og
  • TITAN
  • SingerPretrain
  • OV2
  • KLM 4.1
  • DMR
  • Rigel
  • Nanashi

[vocal mortar]

Ayo? @terse crater level 26 !!!

[visual beacon]

https://tenor.com/view/breaking-bad-patty-cake-gif-25698196

Rigel and Nanashi fighting for worst place lol

[terse crater]

xD

[visual beacon]

KLM 4.1 added
Sample:

[barren zealot]

Idk what this means at all

[haughty panther]

*and Mel roformer

[real flicker]

so my sigma pretrain doesnt exist 🙄

[visual beacon]

I tested the newest ones, but since you said it, I'll do it with yours too lol

[real flicker]

someone is jealous

because i dont use tdeesser

[haughty panther]

without looking the spectograms and only using my ears, for me the original sounds better

i order them from best to bad based in my opinion and ears:

1. original pretrain
2. titan
3. klm 4.1
4. ov2
5. nanashi
6. rigel

[visual beacon]

How is Nanashi above Ov2??? That's the worst pretrain by far

[haughty panther]

fixed lol

heard again to ensure fair opinion and yea, ov2 superior than nanashi (no offense)

[visual beacon]

np, I think he's awful

[haughty panther]

probably

[edgy tulip]

Someone is jealous

Because nobody uses singerpretrain

[barren zealot]

I've never even heard of it

[edgy tulip]

Exacrly

[deep kite]

If you want to accurately test a pretrained model, train new model with each pretrained model for just one epoch using a 1~3 second sample, then test with the resulting model.

[deep kite]

the spectrum I showed to Fitz Roy last time wasn’t a spectrogram of the inferred model but the original spectrogram of the singer's voice recorded by the microphone. so, the harmonics weren’t wrong; they were part of the original recording. this isn’t so much about your judgment being wrong but rather that human eyes and experience are subjective, which leads to issues like this. especially when you set the answer and look for it, you end up staying within the framework you’re already familiar with.

The fact that a pretrained model is trained on a lot of data doesn’t always translate into a purely positive thing. It can highlight weaknesses you might want to hide or amplify incorrect sounds that you accidentally included in the training data.

most data-cleaning processes differ because each method and habit varies. Everyone has different habits, and when similar sounds are collected, it implies there’s a domain of sound that the pretrained model can't handle.

for instance, the breathing sounds of real singers are quite rough because they need to take in large amounts of air in a short period. These sounds can be quite distracting. While they’re usually removed during mixing, I was surprised to find that these breathing sounds didn’t appear at all in the model trained just recently. However, that meant that the breathing sounds originally present in my dataset weren’t being generated at all by the model, and I realized that this wasn’t purely a benefit.

what you should know is that the more extensive the pretrained model’s data is, the broader the variety of sounds the model can infer. A broader variety means that even sounds you might not want to hear can be inferred.

If you really want to know the answer, I recommend training your model without using a pretrained model. While it may take longer, it’s a clear way to monitor exactly what sounds are present in your dataset without the intervention of the pretrained model.

and you’ll quickly spot issues with your own dataset that you might not have noticed before.

of course… there could be technical challenges that make it more difficult, which isn’t your fault, but you’ll understand once you try.

[real flicker]

Like most of other pretrains

Lol

[real flicker]

My pretrain is longer on this server than u 😭

[visual beacon]

Hmmmmm I see what you mean, and I really hadn't stopped to think about it... But I wanted to test the pretrains on how most people will use it

Singer pretrain added
Sample:

[junior walrus]

We can train any dataset length without a pretrain? Or we can only train big datasets (over 4 hours) without a pretrain?

[visual beacon]

I'm testing it now, 200 epochs and the sound is garbage

I'll train more

[junior walrus]

Makes sense, also, wouldn’t be better to use a lower batch if you train without a pretrain?

[visual beacon]

maybe so, but I've already started training at 16, so I'll just keep going

[static storm]

Broader variety sounds won't be replicated if u will not consider them in the contentvec. Breathing is bad in most cases, because of the contentvec dataset is so monotone

Cvec moment

[static storm]

Nah, u need to fulfill the gaps of pretrained absence. Basically just add more data

[barren zealot]

How long does a dataset need to be for a model trained without a pretrain?

[static storm]

Can't say for sure. But back in past, I trained model with 8 hrs dataset+250ep, it was listenable

[barren zealot]

ooh

how long did it take to train?

[static storm]

Sorry, I don't remember

[visual beacon]

In Kaggle? It would take more than 12 hours

[barren zealot]

damn

I do not use kaggle because there's no guide to the rvc disconnect on there

my friend ported the collab to kaggle but no tutorial on how to use it

[visual beacon]

Hmmmm, I believe the code he made doesn't support two gpus in training, so it would take longer

[barren zealot]

he did tell me he edited it but idk how much, Fitz made it btw

[deep kite]

as all we know, large dataset is better if your dataset is good, If the data is insufficient, many artifacts will occur in various parts. however, the point of my writing is not to create a perfect model without pretraining, but rather to observe the phenomena that arise as you monitor the progress at each epoch while training the model without pretraining. As I mentioned above, if you train a 1-5 second dataset using each pretrained model for just one epoch, it won’t learn much at all, meaning that you can use the pretrained model almost like a regular model. So, if you want to check the state of the pretrained model, try using the method I mentioned above.

another reason I'm suggesting you ppl to train a model without using a pretrained model is, aren't you curious why the 'harmonics' and strange noise patterns that you always mention appear? 🙂 the answer lies there, so give it a try. save every 10 epochs and check the spectrogram of the audio inferred by the model generated at each epoch.

[deep kite]

and one more thing to mention, I think it's time for me to take a break. I've shared a lot and had enjoyable moments with you guys for quite a long time, but now I'm feeling a bit exhausted. since I have all the datasets anyway, I think I'll give it another try later if RVC improves further

[real flicker]

"if"

[deep kite]

[sour ore]

https://tenor.com/view/yapping-yapping-level-today-catastrophic-yapanese-gif-13513208407930173397

(i'm joking just in case)

[sour ore]

[deep kite]

take care guys, I'm gonna miss u all 😛

[sour ore]

What do you mean?

[visual beacon]

Nnnnoooooo

devastated

[real flicker]

let me check spectogram

ok so for my perfect eye singer have better harmonics and mid to high end but og have better details

but im nob in spetogram

[visual beacon]

lol

[real flicker]

JUST TELL ME 😡

[visual beacon]

tell what?

[real flicker]

is this true?

or @grand laurel mr tryhard

[visual beacon]

I think he's asleep now

[grand laurel]

It's better to have a good reproduction of the dynamic range of your studio, so yes the og does to better in that regard than all the others

[real flicker]

So dynamic range reproduction over harmonic clarity?

[grand laurel]

og also has better harmonics....

if the dynamic range is reproduced well, that means that the overall quality is good

[real flicker]

I said that I'm spectrogram noob.....

[grand laurel]

rvcs base weights have the best quality thats all u need to know

[real flicker]

No my pretrain have less noise

[grand laurel]

send me those 2 files

the audio files i mean

cuz all i can see here is worse dynamics

plus urs has more aliasing it seems

[sour ore]

https://tenor.com/view/sonia-dangaronpa-sonia-nevermind-no-that's-wrong-gif-4200040968811139212

No that's wrong.

[real flicker]

There are fewer bright patches which means less noise and the mid and high frequency show clearer separation between noise and harmonics pointing/indicating (idk) to cleaner output

But sadly I'm nob at spectrogram 🥲

[sour ore]

OG sounds kinda better.

[real flicker]

It's just little quieter

And u can clearly hear distortion in 0:3

While mine has more air in vocals

[grand laurel]

Yours sounds less like Ariana....

And the quality difference is huge between the 2

[real flicker]

bro literally i have ariana in the pretrain dataset

[grand laurel]

That doesn't help your pretrain's case...

[real flicker]

ik im just saying

[visual beacon]

lol

[grand laurel]

Your pretrain introduces more compressed frequencies up top too

[real flicker]

Sooooo
singer
pluses :
less noise , snr , more stable low freq
better mid freq
lower amplitude varability
minuses:
compressed high freq
more artifacts (little bit more)
harmonic overlap
reduced dyanic range in high req
less harmonic seperation
og
pluses:
better high freq
wider dynamic range
lower harmonic disortion
more detailed harmonics
minuses:
high noise
less mid freq clairty
low freq representation

still

https://tenor.com/view/hwei-sebastian-int-noob-gold-or-silver-gif-15788893945194952515

[grand laurel]

those patches are just different dynamics

your pretrain just sounds weird

just cuz the high ends lack

so it makes it sound more synthetic

[static storm]

Exactly, what are you expecting from a fine tune of an og pretrained? Especially when pretraineds have same pretraineds: hifigan, and same cvec

[vocal mortar]

Ayo? @static storm level 27 !!!

[real flicker]

yall better calm down im joking but that doesnt explain why og sounds disorted

[grand laurel]

How does it sound distorted

Do you know what distortion is

[real flicker]

yes

i mixed since 2019

u can cleary hear disortion in 0:2

[grand laurel]

i have no clue where u can hear the distortion

some would argue that the singer pretrain hinders the quality

and is more distorted

due to the aliasing up top

which og has noticeably less of

[real flicker]

on the word

faaaaaaaaaaast

[sour ore]

Mate, your singerpretrain sounds kinda more distorted than OG.

[real flicker]

naah yall trippin af

[grand laurel]

thats the dynamic range

urs just doesnt go that high

and they sound pretty robotic

[sour ore]

I agree with that.

[real flicker]

still id win

[sour ore]

No, you don't.

[real flicker]

no

id win

[sour ore]

Well, if you think like that, alright.

🐢 👍

[real flicker]

finally i won

ez

[sour ore]

Litsa, just leave Sztef alone. He's not gonna admit his pretrain got flaws.

I would like to think he's trolling us.

[real flicker]

why do u guys take everything so seriously😭😭😭😭😭😭😭😭😭😭