✨│ai-help

[unique rock]

I always run my models through Mel Band Karaoke and then through that model and my models are perfectly clean.

[knotty moth]

it's pretty good but can't remove delay echoes

[unique rock]

I know, to finish eliminating the echoes, I use the normal UVR De Echo at maximum

[latent cypress]

same here, but I did it in the opposite order

[knotty moth]

I found the aggressive one is better

[unique rock]

did it give you a better result?

[unique rock]

which ?

[latent cypress]

maybe you can try melband dereverb mono by anvuew or uvr de echo, but uvr de echo has some cutoff output

[unique rock]

oh, any modifications to his aggression or are you leaving it intact?

[latent cypress]

sometimes

[latent cypress]

yeah the aggressive one did it better imo

[knotty moth]

dereverb mono is good for room reverb in recordings, though it may need to apply RX 11 dialogue isolate

unlike the latter, the former doesn't damage breaths

[unique rock]

The way I do it usually removes the back notes and adlibs well, although sometimes it eats up the high notes too. It's quite clean. I should try your method.

[latent cypress]

yeah I usually use anvuew's dereverb after extracting the vocals, if there are still background vocals left after dereverb I usually go to uvronline (xminus) to remove them using melband karaoke or uvr bve v2

but if the background vocals still not removed, then i didn't add the audio to my dataset

[unique rock]

Removing the back and leads from Billie Eilish's songs is very difficult, but that method helped me a lot and my voice is very clean. Now I'll try your method and see how it goes.

[latent cypress]

most of billie eilish's songs is available in dolby atmos, maybe the stems can help you to get a clean lead vocals

[unique rock]

True, but I don't know why the vcoes in DA are low.

[latent cypress]

which song?

[unique rock]

all of them, meaning the volume is generally low and it is necessary to turn it up, that is where quality comes into play.

[latent cypress]

have you open it in audacity?

[unique rock]

Yeah

[latent cypress]

just normalize the volume or use amplify

[unique rock]

I'll try, thanks for the advice

[red kayak]

thats a breath

[brittle wing]

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[buoyant frost]

Whenever I try to use Inference tool on Applio, and press Convert - it just says processing and the amount of seconds next to it - but it never completes it, just displays "Error" - I haven't gotten it to work a single time

How long does it usually take to convert a 2 min audio file?

[gleaming kindle]

pitch extraction is taking forever, is that normal?

thread:7, f0ing, Hop-Length:64: 100%|█████████████████████████████████████████████████| 22/22 [54:02<00:00, 147.36s/it]
thread:3, f0ing, Hop-Length:64: 100%|█████████████████████████████████████████████████| 23/23 [56:09<00:00, 146.52s/it]
thread:10, f0ing, Hop-Length:64: 100%|██████████████████████████████████████████████| 22/22 [1:46:29<00:00, 290.43s/it]
thread:1, f0ing, Hop-Length:64: 83%|██████████████████████████████████████▊ | 19/23 [1:46:12<35:03, 525.98s/it]
thread:3, f0ing, Hop-Length:64: 100%|██████████████████████████████████████████████████| 23/23 [56:03<00:00, 65.50s/it]
thread:4, f0ing, Hop-Length:64: 64%|███████████████████████████████▏ | 14/22 [55:19<36:17, 272.20s/it]
thread:4, f0ing, Hop-Length:64: 68%|████████████████████████████████ | 15/22 [1:08:59<50:59, 437.06s/it]
thread:5, f0ing, Hop-Length:64: 73%|████████████████████████████████▋ | 16/22 [1:43:21<1:21:19, 813.17s/it]
thread:6, f0ing, Hop-Length:64: 50%|██████████████████████▌ | 11/22 [1:46:07<2:48:07, 917.04s/it]
thread:9, f0ing, Hop-Length:64: 32%|███████████████▎ | 7/22 [37:58<1:40:01, 400.07s/it]
thread:8, f0ing, Hop-Length:64: 64%|████████████████████████████▋ | 14/22 [1:35:28<1:47:45, 808.22s/it]
thread:9, f0ing, Hop-Length:64: 36%|████████████████▎ | 8/22 [1:42:16<5:47:19, 1488.57s/it]
thread:10, f0ing, Hop-Length:64: 100%|██████████████████████████████████████████████| 22/22 [1:46:13<00:00, 124.93s/it]
thread:11, f0ing, Hop-Length:64: 68%|█████████████████████████████▉ | 15/22 [1:43:07<1:24:01, 720.25s/it]
thread:12, f0ing, Hop-Length:64: 68%|█████████████████████████████▉ | 15/22 [1:43:07<1:00:33, 519.13s/it]
thread:13, f0ing, Hop-Length:64: 59%|██████████████████████████ | 13/22 [1:46:00<2:09:05, 860.64s/it]

it has been 3 hours

[knotty moth]

please check some error messages in the console window

[knotty moth]

processing it in cpu mode is supposed to be like that

[gleaming kindle]

I had it use my gpu,,

[knotty moth]

RTX 2060 shouldnt be insanely slow

and it is the lowest recommended gpu for training

[gleaming kindle]

im using a 4060 laptop

i decided to restart it

[knotty moth]

be sure to check gpu usage during the process

[woeful crow]

hi i cant run applio colab and i get this message

ERROR: Failed to launch TensorBoard (exited with 1).
Contents of stderr:
Traceback (most recent call last):
File "/usr/local/bin/tensorboard", line 4, in <module>
from tensorboard.main import run_main
ModuleNotFoundError: No module named 'tensorboard'
Traceback (most recent call last):
File "/content/program_ml/app.py", line 1, in <module>
import gradio as gr
ModuleNotFoundError: No module named 'gradio'

[buoyant frost]

how do i do that

[gleaming kindle]

restarting it seemed to fix it pitch extraction for some reason, its almost instant now

[knotty moth]

ask google or chatgpt

[grim bay]

I would say yes

But not sure if it is

[blazing solar]

Can yall help me this?

Hello?

[low shard]

Applio colab is broken for now @blazing solar @woeful crow

This is a Google colab/uv issue

[woeful crow]

thanks

[low shard]

Yw, use an alternative for now

[blazing solar]

What kind?

[low shard]

Train (make) RVC Models on cloud:

1. Prepare the Dataset
2. Setup RVC:
   Choose a cloud way to use RVC,

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
• Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
  • Mainline (UI)
  • Applio by Vidal (UI)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Applio (UI)
  • Mainline (UI, No guide as of right now)

3. Be sure to know about the tensorboard

Google Colab = Easier but risk of getting disconnected
Kaggle = Harder but way more gpu time
If you are looking for the easiest way and for free, is using https://weights.com/ which ofc uses RVC

RVC Inference (use models) on pre-recorded audio on Cloud

You can use either:

• Weights.com: Easiest Possible Ever Automatic
• Ilaria RVC Zero: Fastest free on cloud
• Applio UI Colab: RVC Fork with some extra features like TTS
• RVC AI Cover Maker UI: Automatically Separates the Vocals and Instrumentals, converts the voice and mixes them back

Or do you have a PC? Did you try checking it's GPU

[latent cypress]

Is it fine to add the same song but in a different language to my dataset?

[latent cypress]

Is there any documentation on this?

[white bough]

Yes, the model will learn new sounds that come from different languages

[latent cypress]

ah okay, thanks for answering my question

[white bough]

For colab, atm you can replace all the instances of "!uv pip" in the install cell by "!pip" and it should work, although it might take longer

[brittle wing]

Hi, may I ask what was the way to clone the main repository into applio no UI colab so I can see the newer average graphs ?

[brittle wing]

-guides

[simple ore]

in the clone cell remove --branch parameter

[brittle wing]

Only this just removing it will do the thing?
Also where's the guide for this colab, I lost it

[unique rock]

What aggression do I put on the UVR Aggressive? 0.5 ? or 1.0?

[full tundra]

What rvc I can use to create and train voice model?

[odd shale]

You can use any version, but we mostly recommend Applio.

[surreal latch]

One question, have the creators said anything about this?

[knotty moth]

gatekept

[unique rock]

that ?

[brittle wing]

what is the link for the browser rpc

like the one u can use in browser

[knotty moth]

what rpc?

[brittle wing]

its called like liana rpc or sum

i remember it was a website

[knotty moth]

sounds like ur dyslexic

https://huggingface.co/spaces/TheStinger/Ilaria_RVC ?

[knotty moth]

https://tenor.com/view/huh-cat-gif-26460616

[brittle wing]

sry

[hallow thistle]

The correct name is Ilaria RVC. I think you just misheard it somewhere.

[knotty moth]

nah just being dyslexic

[hallow thistle]

Too crazy.

[hallow thistle]

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[hallow thistle]

This is the correct command for showing guide links to RVC and Applio.

[olive hill]

What is this error on Kaggle

[hallow thistle]

Which Kaggle notebook is this?

[olive hill]

The one mentioned in Applio Kaggle Doc

https://www.kaggle.com/code/deiant/applio

[olive hill]

can u help?

[hallow thistle]

Try delete the notebook you created for yourself and then create the new one from start.

[olive hill]

Nope , tried same issue

[olive hill]

@hallow thistle Please help

[knotty moth]

try doing this #✨│ai-help message

[low shard]

they reported it to uv and google

[serene horizon]

So what is the cause of static/crackle in silence gaps in resulting file from inference?

I convert the same piece of audio, and the model I created in Applio has it, while the model I created in RVC Mainline doesn’t have it.

[knotty moth]

that happens when changing envelope unlike mainline inference

[serene horizon]

Envelope? Im doing the inference in the same program, all I change is the .pth file between Applio created and Mainline created.

[knotty moth]

seems like you didnt use this slicing method

https://cdn.discordapp.com/attachments/1159290139609137264/1356847850146762824/image.png?ex=67f74908&is=67f5f788&hm=dc8923b1bb1b35da996c081c781efbeba432bc535e975db21b3b18b0214da1f5&

[serene horizon]

I used exactly the same files for model creation in Mainline and Applio.

They are already edited to remove excessive silence.

[knotty moth]

both have different slicing method, thats why

so you might want to fall back to the old labeling method for use in mainline

[serene horizon]

The Mainline model is better, less noise.

[knotty moth]

you need better denoising perhaps and dont truncate silences too short

[serene horizon]

In my audio for model training? There is no noise. And only enough silence for breathes.

Also, Applio has more “electronic” noise on actual voice than Mainline model. Especially with “SSS” sounds.

[knotty moth]

can't really say that, the result may vary on different epochs

some training configurations are actually different tho

[serene horizon]

What is strange is I tried conversion in the Weights Replay app, and it works well, but it doesn’t ask for index file.

[latent cypress]

is applio main automatically set to fp32?

[analog obsidian]

latest main branch? yep, fp16 got removed

in the latest compiled version fp16 is enabled by default

[latent cypress]

ah i see, i was confused because i didn't find any option to change precision

[latent cypress]

i downloaded the apllio from code

[analog obsidian]

nice, fp16 doesn't exists there

everything is fp32 now

[latent cypress]

okay, thanks

[latent cypress]

is using checkpointing really that slow? it's been 5 minutes after i click start training, and the cmd didn't show anything

[analog obsidian]

uh no?
it's very slow for 48k tho and i haven't tried 40k
for 32k the speed is ok

maybe you've got out of vram even with checkpointing enabled

i know that 48k requires way more vram to train models

40k should also require a bit more vram than 32k

[golden walrus]

guys, 1 question, is it possible to train a model that can be smooth in my language?

[latent cypress]

oh really? tbh i can do 8 batch size without checkpointing on codename fork, but i'm just worried about my pc if i continue the training

[analog obsidian]

yes because codename's fork still has the old inplace flag

that saves vram

but it got removed in applio because it was causing problems when training from scratch

[eager siren]

Is there any recommended cloud collab for training?"

[blazing solar]

Im using a mobile here and i cant access to any of these

[latent cypress]

oh, lemme set the batch size to 6

[analog obsidian]

i've heard that rvc sometimes adds an american accent in models, but I'm not sure about that, I personally haven't had any problems with this.

however you can use the index to force the accent of the dataset in the results

[golden walrus]

somehow i train it with my language data but end up sound like a British

[analog obsidian]

yea you're not alone in that, idk why it happens

never had that issue personally

[latent cypress]

you're right, i just set the batch size to 6 and it worked

i guess i have to try kaggle sometimes

[analog obsidian]

yes training is faster because there's a new flag named "benchmarking" that speeds up the training

[latent cypress]

oh wow, it's been a while since I've followed the updates

[analog obsidian]

but noobies told me that the speed boost was more noticeable while using fp16, and that fp32 negated the speed gains

[low shard]

wdym

[latent cypress]

it''s very noticeable here, it's like I was training with fp16 back when I was still using rvc disconnected colab

[analog obsidian]

for me it's also very fast lol, specially when training 32k

[latent cypress]

for some reason i doubted that my applio is using fp32 lol

[analog obsidian]

you can have fun and experiment with it if you want

go to rvc/train/train.py

find

torch.backends.cudnn.benchmark = True

set it to false to disable the speed gain

i have not compared how slow the training is with that setting turned off xd

[latent cypress]

is pretrained 32k is used if your dataset is mostly lossy?

[analog obsidian]

is used if your dataset has a 16k cutoff

there's no point to train 48k with a 16k cutoff

slower training + high vram usage for basically the same results as if you were training using 32k

[latent cypress]

what if my dataset is mostly lossless, but there is one audio has 16k cutoff, is that okay to use 40k pretrain?

[analog obsidian]

if your dataset has one audio with a 16k cutoff but the rest isnt you have two options:
resample the entire set to 32k
remove that specific 16k audio and train only using the lossless audio

[latent cypress]

the audio source is actually lossless, but after removing the vocals and reverb, some echoes still remained. then i removed them using a de-echo model, which caused this cutoff.

[analog obsidian]

oh damn
you can either train 40k or 32k

i would use 32k

the dataset now has a 17k cutoff

and also is damaged due to separation models being used

[latent cypress]

trying with 40k rn

[latent cypress]

the frequency?

[analog obsidian]

you won't notice a difference

the dataset doesn't reach 20k nor 18k

17k is pretty much 32k territory

[latent cypress]

is the sound detail still good at 32k? haven't tried 32k until now

[analog obsidian]

rvc doesn't care about sample rate

you'll notice a difference if the dataset is actually true 48k/40k

but if the dataset is 32k instead, then no, training that in 48k/40k doesn't give you better results

[analog obsidian]

it's just going to copy your frequency cutoff

[latent cypress]

i just thought 32k is less detailed than 40k all the time lol

[analog obsidian]

it is but only hardcore audiophiles can hear the difference lol

40k and 48k are more clear than 32k

but again, only if the audios are true 40k or 48k

17k from separation models is not true 40k nor 48k

[latent cypress]

ah okay okay

[analog obsidian]

if your dataset had a frequency cutoff of 24k you can downsample it to 32k and compare them
see if you can hear some difference lol

[analog obsidian]

i mean the original audios, with all of the reverb and sheet

if you still have them, try to convert one of the audios to 32k

and try to hear if there is a big difference

i did it rn and i cant hear the difference

[latent cypress]

me neither lol

[analog obsidian]

see? don't be scared of 32k

48k has more fidelity than 32k, buuut, only hardcore audiophiles can hear the difference

[latent cypress]

i guess i don't have to worry about mix lossless audio and the audio from youtube into my dataset

[analog obsidian]

no wait don't do this

the dataset needs consistent quality

dont mix stuff

[latent cypress]

[analog obsidian]

[latent cypress]

i tried to make a singing model, tbh, my dataset is kinda repetitive if i only use lossless audio, while the singer is hitting high notes on youtube lol

[analog obsidian]

well you can experiment and try training your current dataset, but it is a known fact that training from different sources can lead to bad results

[latent cypress]

I always do this when training the model, but for some reason people say my model quality is really good, i even got ai master role on some server

[analog obsidian]

im just telling you what i know
if u want to continue training like that, do it lols

at the end what matters is if you like how your models sound

[latent cypress]

no, i mean i feel so bad for people who always use my model, of course i will change how i do my training now, thanks for giving me some advises

and now I feel like I don't deserve the ai master role on that server

[analog obsidian]

ai is random
i also make mistakes while training models

this thing can be really confusing sometimes lol

im still learning more about rvc

it's all trial and error here

[latent cypress]

yeah, you're right

[grim bay]

@viscid moss Why does only UVR-Deecho-Dereverb works? The rest models in Dereverb tab are just giving error

[viscid moss]

I'm working on that

[opal kelp]

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[knotty moth]

no, it will only fry if your pc has bad airflow and bad thermal paste applied on the cpu/gpu

[storm holly]

not sure it works anymore

[knotty moth]

details ≠ additional high ends by going higher than 32k

[knotty moth]

2 min dataset with consistent & fine quality could be better than the longer one with such mixed sources

[knotty moth]

it has been fixed

[golden walrus]

guys, is there any guide on how to use UVR UI ?

i've met some error but i have no idea where i went wrong

like this

or

[knotty moth]

not quite there

[golden walrus]

i don't get it, why error, what file

[knotty moth]

check the file path name

[golden walrus]

is it environment path or

the name of the voice file ?

[knotty moth]

D:\path\to\your\file.wav

[storm holly]

wait i need some kind of ngrok token for applio colab?

[golden walrus]

is this the voice path ?

or the uvr path lah

[knotty moth]

if I were you, I would have known where is your wav file you'd want to process

[golden walrus]

ahhhhhhhh

i thought uvr just need to throw in that file and done

because i don't see any output path

or input path in the ui

[knotty moth]

try another path like D:\

I mean like C:\yourfolder\your.wav

also you can try first pre-convert the mp3 file to wav using audacity

[golden walrus]

roger that

[knotty moth]

[latent cypress]

yeah, that's why i stick with batch size 6

[hot dome]

chat can you chek it's minus 1 minutes i cook https://www.youtube.com/watch?v=r_bGpHbdSy0&ab_channel=Airashi

[golden walrus]

sire, i failed you. it still doesn't work

i think i have to download the model and import it in the file somewhere

[limpid birch]

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[latent cypress]

do you think it's overtrained after 8.7k steps?

[viscid moss]

Still having issues?

if still... can u show me this part of the error?

My DMs are also open for help

I think the problem is the installation path, it's directly in C:\ Try installing it in Downloads

[unborn stirrup]

Is anybody else having issues with Mainline? I keep getting error "ERR_NGROK_8012".

[wanton valley]

same

from the #📰│dev-updates it seems there are still issues with the damn colab so i guess we wait

[tropic garden]

Me too! I get the same error on Kaggle too.

[simple ore]

if you're using UI colab you need a way to connect a browser to colab, it can be done with a public gradio link or with ngrok tunnel

[simple ore]

are you training without a pretrain or something? 45 g total is very high

[analog obsidian]

old g/total graph

[knotty moth]

https://ngrok.com/docs/errors/err_ngrok_8012/
tl;dr possibly incorrect/screwed code cell to run the webui through ngrok, or the previous tunnel session wasn't successfully stopped so the port to connect was in use

[viscid moss]

@left flame add those lines at the beggining of the cell

[solemn walrus]

.

[keen island]

hello, does the rvc gui also exist with RMVPE

[knotty moth]

nuh uh

-gui

[patent trellis]

RVC Easy GUI

This program is pretty abandoned and old, don't use it.

[hallow thistle]

-gui

[patent trellis]

RVC Easy GUI

This program is pretty abandoned and old, don't use it.

[knotty moth]

^suggestion: the bot should show a link to the latest rvc or applio: https://docs.aihub.gg/essentials/whats-rvc/#forks

[hallow thistle]

Don't expect RVC-GUI to get updated with rmvpe model; this particular RVC fork is too old. Use a better RVC program like Applio instead.

[keen island]

Oh okay thanks, I'm not up to date, which is the easiest?

[brittle wing]

Applio no UI doesn't wat to preprocess dataset

[keen island]

I actually found the rvc gui quite good, audio in, audio out in the same folder

[knotty moth]

why good if there's no rmvpe?

[keen island]

That's exactly why I asked if the gui is also available with rmvpe, well I just have to switch

[latent cypress]

im using klm hifigan final

[knotty moth]

rmvpe is better than others in most cases

[hallow thistle]

Applio. It's more up to date.

[knotty moth]

except that it may struggle on inferring polyphonic/choruses

[keen island]

I have already loaded trained voice models into rmvpe. Which folder do they have to go into in apollio?

[hallow thistle]

rmvpe isn't RVC program. It's a pitch extraction model used in RVC.

[brittle wing]

Applio no UI doesn't want to preprocess dataset

Traceback (most recent call last):
File "/content/Applio/core.py", line 15, in <module>
from rvc.lib.tools.prerequisites_download import prequisites_download_pipeline
File "/content/Applio/rvc/lib/tools/prerequisites_download.py", line 3, in <module>
from tqdm import tqdm
ModuleNotFoundError: No module named 'tqdm'
@simple ore also in the "download models" sub cell

[latent cypress]

the voice is also weird, idk what is called, overlapping?

[knotty moth]

the installation seems screwed

[brittle wing]

Hm well then?

What do I use for training

[knotty moth]

host BackEnd?

are you doing this shit for rvc?

[brittle wing]

What do I use for training is there something that works?

[knotty moth]

does it use rvc or nah?

do you know what purpose of this channel so that you're asking it here?

anyway you could ask chatgpt or claude

[weak cipher]

sr bro

I will delete the message

[low shard]

what’s your pc gpu and what do you want to do?

[brittle wing]

Jokes on me...
Actually I'm a phone user Applio no UI used to work fine, now what happened

What happened

[low shard]

yeah the issue is that uv and google colab are having issues with eachother, applio reported this to both, so all we have to do is wait for uv or google colab to fix it

you can use alternatives

Train (make) RVC Models on cloud:

1. Prepare the Dataset
2. Setup RVC:
   Choose a cloud way to use RVC,

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
• Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
  • Mainline (UI)
  • Applio by Vidal (UI)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Applio (UI)
  • Mainline (UI, No guide as of right now)

3. Be sure to know about the tensorboard

Google Colab = Easier but risk of getting disconnected
Kaggle = Harder but way more gpu time
If you are looking for the easiest way and for free, is using https://weights.com/ which ofc uses RVC

RVC Inference (use models) on pre-recorded audio on Cloud

You can use either:

• Weights.com: Easiest Possible Ever Automatic
• Ilaria RVC Zero: Fastest free on cloud
• Applio UI Colab: RVC Fork with some extra features like TTS
• RVC AI Cover Maker UI: Automatically Separates the Vocals and Instrumentals, converts the voice and mixes them back

ofc don’t mind about applio since it doesn’t work right now

[brittle wing]

Mainline?

[low shard]

It happens often that cloud breakes, especially google colab

because notebooks devs need to adapt to every cloud site update and bug, which is tiring

Or the other option would be buying a good pc, but not everyone can afford that

[brittle wing]

Hmm I just want a working Colab

[low shard]

mainline is the og rvc project yes, it doesn’t have the same perks like easy ui and tts or some adjustments, but it should work

[brittle wing]

Yes but it doesn't have option to input pretrains.

The Colab

Mainline

[low shard]

I mean it’s on cloud and google decides what to update on those pc, cloud is not as stable as local (running it on your own good pc)

google colabs break basically weekly, check also #📰│dev-updates

I’m also checking right now if mainline colab works

[brittle wing]

Okay

[low shard]

because the same creator of the mainline colab, hina, is busy and some others of their colabs are broken like wokada, so i’m not sure if mainline is broken too

[brittle wing]

I hope mainline works at least but I find it difficult to use

[low shard]

Mainline colab does

[brittle wing]

Now?

[low shard]

i mean, if you really need to use applio,u could try the lightning.ai site version, which gives only 22 hours monthly of gpu and is harder to setup and needs a phone number verification

else, the only other working way to use applio is to buy a good pc

because applio kaggle is broken too

or you could try other RVCs like mainline or rvc ai cover maker ui or ilaria rvc

[low shard]

there’s a whole cell in the hina modified mainline colab about pretrains

[brittle wing]

Uh that's perfect

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[latent cypress]

or pay the model master who opened the commission

[low shard]

nevermind, just tried it, mainline colab is broken too

[brittle wing]

Yes but the pretrain I wanna use isn't listed in

[low shard]

so yeah use other alternatives

[brittle wing]

Whaaat

I can't train?

Applio UI colab?

[low shard]

broken too

cloud is unstable asf

[brittle wing]

When will. It get fixed

I have a model to train

[low shard]

who knows,

[brittle wing]

...

[low shard]

for applio i told you we need uv and google colab to fix it, it’s their problem that causes issues for applio

it’s not an applio issue, it’s related to another thing that applio uses for cloud

[low shard]

your other option is weights.com

[brittle wing]

I understand but when will it get fixed

[low shard]

or paying a model master commission, or wait for someone o do your free model request

[brittle wing]

Yes but it doesn't let me use pretrains

I know that

[low shard]

no one knows, there isn’t a specific date

[brittle wing]

Weights doesn't allow pretrain input

[low shard]

i’m not sure if weights uses its own pretrain, tho u could suggest that feature in their server

[latent cypress]

medleyvox really ruined the frequency, and I didn't realize I was training the model with this audio

[low shard]

@brittle wing your only options are:

• use weights
• use other cloud alternatives, like mainline or applio on lightning ai
• buy a good pc
• ask a model master or maker to make it
• wait for fixes
• pay for cloud

[brittle wing]

I wanna train it myself I have actually trained models but the notebook doesn't work

[low shard]

i understand, i told you the only options available

[brittle wing]

Applio colab?

[low shard]

i just told you it’s broken

no other ways than the ones i said

unless someone makes a script fix

[brittle wing]

...

I see no inference, no training

[knotty moth]

the kaggle notebook is another option

[brittle wing]

Link?

[latent cypress]

https://www.kaggle.com/code/deiant/applio

https://docs.aihub.gg/rvc/cloud/applio-kaggle/ (how to setup)

[low shard]

that’s broken too iirc

[brittle wing]

This is so complicated

[brittle wing]

It's complicated

E complicato

[low shard]

the interface is harder than colab but ujust gotta read the guide

[latent cypress]

no need to worry then, the notebook is also broken

https://cdn.discordapp.com/attachments/1244363710986715247/1270699138581008405/ag.gif

[brittle wing]

✅

[latent cypress]

so the only works cloud notebook is only lighting ai?

[low shard]

For rvc yes

[brittle wing]

Idk

[brittle wing]

Link?

[latent cypress]

sorry wrong reply

[low shard]

-lightning

[patent trellis]

⚡ Lightning AI Studios

Lightning AI is a Cloud (Remote Good PC) Service. The Free plan provides up to 22 hours of monthly usage.

• **Applio**

by Shirou
Lightning AI

• **RVC-WebUI**

by Shirou Lightning AI

• **FaceFusion UI**

by Nick088
Lightning AI

• **UVR5 UI**

by Eddy
Lightning AI

• **UVR5 NO UI**

by Eddy
Lightning AI

[low shard]

the first 2 should work

[brittle wing]

Does Ilaria rvc on hugginface work?

[low shard]

yes but only inference

[brittle wing]

For inference

[brittle wing]

That's what I'm asking

Is it the same result as Applio?

How do I use the lighting notebook

[latent cypress]

https://docs.aihub.gg/rvc/cloud/applio-lighting-ai/

[brittle wing]

How do I use this

Wait it's similar to colab

[latent cypress]

idk, i haven't tried it. but we got the docs to know how to use it

[brittle wing]

Okay does it allow pretrains

[low shard]

you asked for training

traiming = make models

iference = use them

[brittle wing]

Yes...

[low shard]

it’s the same as applio, just different cloud site, on lightning ai the interface is a bit harder, needs a phone number verification which could take some days for them to verify u, like 2 or 3, then also gives u a max 22 hours of gpu monthly
But the program is the same

[brittle wing]

But this isn't okay can you train the model.

[low shard]

it is okay, it’s for free, you can’t expect free gpus 24/7 for free

[brittle wing]

I'm waiting for them to verify me

[low shard]

gpus are expensive

[brittle wing]

Colab was better

[low shard]

and yeah you can train the models on lightning.ai applio

[brittle wing]

22 hours a month aren't enough

[low shard]

colab igves more gpu time and is easier to use, but its broken

[brittle wing]

When will the fix come...

[low shard]

i can’t do anything about that

[brittle wing]

I know

[low shard]

you can pay for more gpu time on lightning.ai

[brittle wing]

I'm just saying

[low shard]

just like you can for colab too

[low shard]

don’t need to ask it again, we don’t know

[brittle wing]

I don't wanna pay Money for such things

I need colab back

[low shard]

you can’t expect to get the easiest service 24/7 all for free

[brittle wing]

Who do I ask

[low shard]

AI is expensive asf

[low shard]

no one

have patience

[latent cypress]

I always thought buying a good pc (at least rtx 4060) is an easy thing in other countries. considering the salary from their job is at least bigger than mine which is only 2.6M IDR ($162.50) per month.

[brittle wing]

How do i fix it

[low shard]

i’m glad I got an rtx 4060 ti16gb pc

[brittle wing]

Is it s python script problem or a general one.

[tropic garden]

We buy Macs instead and get stuck. 😥

[low shard]

Unless you know coding and are able to find a workaround, all you can do is wait for google colab and uv to fix it, it’s a problem of the site hosted machines and of the python dependencies

[latent cypress]

[simple sundial]

hey guys whats up

[brittle wing]

I give up

[low shard]

Macs are perfect for people who want easy to use pcs
But AI is mostly good on nvidia gpus, since they are more powerful and got more support

[brittle wing]

...

[low shard]

I mean you can just use the options I told you

[simple sundial]

Is this the right help channel for the voice-changer?

[brittle wing]

Weights and...?

[low shard]

I’m talking about those ones

[brittle wing]

Weights isn't sophisticated enough no pretrains

WTF

[low shard]

which type?
RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models

Wokada = uses RVC for realtime inference

[latent cypress]

tbh kinda regret getting a PC with an rtx 4060 8gb. now I can’t even train my models with a batch size 8

[low shard]

Its meant for simple to use RVC usage yeah

[brittle wing]

I know that

[simple sundial]

RVC, yup. I'm using a realtime voice changer client and wanted some advice on models since I don't fully understand something about it

[low shard]

i mean you can use batch size 4

[low shard]

rvc isn’t a realtime voice changer

rvc is for pre-recorded audios and model training

[simple sundial]

does it not stand for that?

ahhh

[low shard]

Did you read the whole text i sent?

please read it

[simple sundial]

nvm, I see. some words I didn't understand just flew by me

[latent cypress]

it's different hehe

[low shard]

i’m guessing u want a realtime voice changer, so wokada, ** be sure to not use yt tuts**, elaborate ur issue in #🔍│help-w-okada

[simple sundial]

thanks then!

[latent cypress]

is the graph stable on batch size 4?

[low shard]

ipinged u in it

[brittle wing]

Can I dm you so you train the model...idk anymore

[latent cypress]

I know it will improve accuracy, but codename fork says the graph won't be as stable as batch size 8 (iirc)

[latent cypress]

i also trained my model now

[brittle wing]

I know I saw that

@low shard idk maybe someone found a fix

[low shard]

Yeah

[low shard]

you need to have patience unfortunately

[brittle wing]

Look at this...

[low shard]

No that’s an old fix that already got added, the issue is related to something else now

all you need is wait or use the alternatives

i don’t know how to tell you but there’s no other way

[brittle wing]

WHAAT

[low shard]

you don’t have other options

[low shard]

yup

colabs break weekly

cloud in general breaks very very easily

[latent cypress]

like the fix was made in March, I'm not sure that's the way to fix clouds for now.

[brittle wing]

Exactly

[low shard]

there isn’t as of right now

it’s another type of issue

[brittle wing]

But how some other colabs still work

[low shard]

because they use different code

i understand you need colab, but there’s no other way

Please just follow the options i told you

[brittle wing]

What can I do to fix the bug

[low shard]

keeping to ask won’t fix anything

[low shard]

nothing unless u work for UV and google and know coding

[brittle wing]

The fix I sent was uv fix

[low shard]

no

it was for pydantic

look

[brittle wing]

Since when are the colabs malfunctioning

[brittle wing]

Ah yes

[low shard]

Some from momths, some from less, like applio since some days

[brittle wing]

I see

[low shard]

It doesn’t matter if you keep asking, the answer will be always the same, cloud gets broken weekly

hopefully things get fixed soon

[latent cypress]

if it were that easy to fix, the devs would've handled it by now. The fact that it’s still broken probably means even they aren’t sure how to fix it.

then it's useless to tell you how to fix it

[brittle wing]

Exactly

[low shard]

the issue isn’t even applio’s fault, it’s that uv is having issues on google colab, which is used for installing applio requirements

[brittle wing]

What is uv

[low shard]

Python library

[simple ore]

uv is a new requirements installer

the workaround is here https://github.com/IAHispano/Applio/issues/1025

[low shard]

Ima announce that

[brittle wing]

And?

[brittle wing]

Will this help with the fix answer please

[golden walrus]

guys, is the overtraining detection working ?

[low shard]

Yes if you follow it

You need to modify the code like in the issue discussion

I mean that manually for yourself

[latent cypress]

i have so many audio with varying frequencies like this on my dataset, should I delete the audio that has cutoffs, or keep it and resample everything to 32k, then train it using a 32k pretrain?

[simple ore]

and what

[brittle wing]

Nice but will it work in applio no UI also how do I input the dataset on there as a zip or a folder with audios?

[latent cypress]

lyery already gave me advice, but I just want to hear other users' opinions

[brittle wing]

If I'm being honest you're thinking too deeply into it.

Or you wanna know everything.

[latent cypress]

i just want to improve my model quality

[brittle wing]

Wow

[latent cypress]

so i need to know how to train properly

[brittle wing]

You already know

I think

[latent cypress]

not really

[brittle wing]

Okay just a question how should be my dataset be In applio as a zip file or a folder with audios?

Cause sometimes it's different

[latent cypress]

idk if it's in the cloud, but on local I just need to copy my datasets folder to "assets >
datasets"

[brittle wing]

The colab Noobies sent a fix there

[latent cypress]

what fix?

[brittle wing]

https://github.com/IAHispano/Applio/issues/1025

I know you have trained on Applio colab that's what I'm asking

[brittle wing]

Zip file, right?Not a folder in drive containing audio files

[latent cypress]

yeah extract the folder with the audio inside into zip, so you will get jhope.zip\jhope

[brittle wing]

Oops how did you know...

[latent cypress]

hahaha

[brittle wing]

Wait what do I do

[latent cypress]

too easy

[brittle wing]

You know?

Ah okay

[latent cypress]

i mean i recognize you

[brittle wing]

But should it be a zip file uploaded into Applio

[latent cypress]

https://tenor.com/view/i-don't-know-idk-idk-about-that-gif-7336250873949261946

never used applio colab before

but on local you don't have to zip it

[brittle wing]

Well where's THE guide.

[latent cypress]

https://docs.aihub.gg/rvc/cloud/applio-colab/#training-

[latent cypress]

everything's here

[brittle wing]

Yes

THANKS 💪

[knotty moth]

useless for those who know how to properly use tensorboard and test checkpoints

[latent cypress]

do you have any screenshot showing that overtraining detector is detecting overtraining?

[brittle wing]

@simple ore IT WORKED

Like a miracle

[latent cypress]

https://cdn.discordapp.com/attachments/767947630403387393/1324675633237983373/bp.gif

[brittle wing]

Yes but I didn't understand about the format zip file or folder with audios.

[brittle wing]

Nvm

I understood

Is "Load a backup" for resuming?

Where's the pretrains tab.

Nvm found in advanced settings

[rocky radish]

Hi guys im kinda a noobie how do i get my macbook voice changer to work

How do i connect the ai voice thing on mac to discord

[brittle wing]

Hmm how do I import dataset in Applio colab

[latent cypress]

is it normal for the loss/g/total starting from the bottom and not from the top?

[brittle wing]

Will merging all the dataset audios work for the Colab?Sorry for bothering

[latent cypress]

idk, i never tried applio collab

[brittle wing]

Well the zip archive didn't work it kept returning errors

[brittle wing]

Still doesn't work

[brittle wing]

I'm just wasting GPU for nothing

[echo drum]

I have a question guys. I heard some voice changers which also go with your depths and highs of your voice without sounding so electronic all of a sudden. Is there a quality difference depending on the model and where can I get a high quality one?

[simple ore]

it is not starting at the bottom. Turn off ignore ouliers, turn off smoothing, you'll see

[brittle wing]

How do I input dataset into Applio colab so it doesn't return errors?

Like as a zip file or audio file folder or a merge of all audios???like I need help.

I just wasted 2-3 hours of GPU just waiting for an answer...

[warm iris]

is there a google colab RVC coverter?

-RVC

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[oak plank]

#📰│dev-updates message

[simple ore]

you make a folder on google drive with wav files

then you provide a path for noUI colab

if you're using UI colab, then you make a dataset and upload wav files

[brittle wing]

No UI doesn't work.

[simple ore]

it does

[brittle wing]

How...

I tried and it don't work

[brittle wing]

Any fix

[simple ore]

#✨│ai-help message

add this at the start of the install cell, run install

[brittle wing]

Will the bug fix you provided fix that one too?

[simple ore]

it is the same, you just need to run it before you run the install

[brittle wing]

I tried & it didn't fix

[lunar perch]

hello all 🤗
im searching a open source or free voice cloning with emotions.
best would be, the emotion reference audio can be a different language.
can you please give me some tipps, im near blind from searching google 😦

i want to dub some old anime to my homelanguage, so i used applio to train some models.
then tts for dub the new audio in my homelanguage. actual im testing f5-tts for the emotional speaking.
but a more "automation" on my work would help me a lot, so i start searching for something
that can help me speed up the process and is not limited to english/chinese language only.
thank you!

[brittle wing]

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[warm iris]

can I get an actual download link without signing up for weights.com?

[oak plank]

Download link of?

[golden walrus]

question, is it possible to stop training mid way ? because i accidentally set 500 epochs

[warm iris]

https://www.weights.com/models/clm739s9s1xntcctc96sjovbg

The computer that sung Daisy Bell

nvm, got it

[pastel oak]

As long as you have save Checkpoints enabled (check if you have G and D files in the models folder) you can stop at the next checkpoint

[golden walrus]

yay, thank you so much

[oak plank]

[latent cypress]

ah okay

[analog obsidian]

use loss avg50, its more accurate than old loss graphs

[latent cypress]

this one?

[analog obsidian]

yes

after some time the model will converge to a value
it'll have a flat line pattern

[latent cypress]

The monitoring method is the same as loss g total, right?

[analog obsidian]

similar
the model then is going to stop going down and its gonna stay in a flat line

[latent cypress]

stay in a flat line means overtraining or what?

[analog obsidian]

means the model is near convergence
its gonna start doing small adjustments, during that there is a risk of the model forgetting the pretrain

[golden walrus]

ohhhm where can i find this one

[analog obsidian]

when you get to that point what you can do is to let the model train for maybe 1 hour more and choose the lowest point and see if you like how it sounds

[latent cypress]

I didn't even set checkpoints to be saved per 1 epoch, kinda worrying about my ssd tbw if I do that

but thanks for explaining

at least i know what to do after this

[latent cypress]

just type "loss" on scalars

[analog obsidian]

well if you dont save every epochs then it's mostly random and you have to hear the epochs until you find the one you like the most

[golden walrus]

i only have these

[latent cypress]

i used applio mainline, idk what are you using now

[golden walrus]

same applio, it's 1 8 3 something

[analog obsidian]

thats the compiled version

still has the old loss graphs

[remote karma]

hello, i downloaded okada but after extracting and opening the exe it opens up cmd but nothing happens after any idea on what I should do?

[golden walrus]

do you mind show me how can i download precompiled version?

[analog obsidian]

they are not useful for choosing epochs, they are very inaccurate
in the case of having those graphs the best thing to do is to listen to the epochs and choose the one you like the most

[golden walrus]

or a way to update

oh, okay, i need to test it out right ? these are just for reference ?

[latent cypress]

yeah, that's what i always do

[analog obsidian]

https://github.com/IAHispano/Applio
and place the applio folder in a NON onedrive location

after extracting it, run this

let it install the env, don't close the cmd until the installation is done

after that, run this and you can train like always

[latent cypress]

ah, that's the reason why I got bad results even though I chose an epoch with low points

[analog obsidian]

mainline graphs only logs the last step

which is most of the time, rvc learning the mute file

[latent cypress]

rvc training always got me feeling nervous

guess i’m just being extra paranoid

[golden walrus]

me too

i'm too new for this field

[analog obsidian]

why?

🦈

[latent cypress]

idk bro, i’ve been training since everyone was still using svc, so I shouldn’t have any reason to be nervous, but somehow, there’s always that bit of worry whenever I do this

i might’ve been training for a long time, but it was never really backed up with the right knowledge on how to do it properly

[jolly pivot]

I am having trouble with the Applio cloud it won't start applio. It will install but it will not start after installing it says error. any fixes

[latent cypress]

https://github.com/IAHispano/Applio/issues/1025

[jolly pivot]

Thank you

[low shard]

always be sure to check #📰│dev-updates

[twilit forge]

if "const" in schema:
TypeError: argument of type 'bool' is not iterable
An error occurred launching Gradio: When localhost is not accessible, a shareable link must be created. Please set share=True or check your proxy settings to allow access to localhost. i keep getting this when running apllio locally

[simple ore]

that requires gradio update as one of its dependent libraries messed it up

if the requirements have gradio >5.0.0, then can go for 5.23.1

[twilit forge]

what bout 5.24

[placid holly]

ah they Merged Voice Changer chat here

[latent cypress]

the time is about to come

[placid holly]

may i ask the link and other stuff directed to W-Okada again, because no pinned message yet for it

[analog obsidian]

yuh, you can try training for a little longer and then start listening the epochs

you have very high values in your graph

thats weird uhh
small datasets usually converge to a value of 35 (or even less than this)

[latent cypress]

okie

[latent cypress]

mine is 22 minutes

[analog obsidian]

you were training without a pretrain?

[analog obsidian]

https://rentry.co/forkvoicechangerguide

[latent cypress]

no, i used klm hifigan final ver

[analog obsidian]

and thats even weirder because klm usually gives me lower values than og pretrain

[analog obsidian]

why are you using overtraining detector 😭 😭 😭

[latent cypress]

because i can't really monitoring the training properly, i thought it's gonna automatically stop my training if it was overtrained😭

[edgy tangle]

Wut

[latent cypress]

should i start it again from the beginning?

[analog obsidian]

no

[analog obsidian]

#✨│ai-help message

ive asked him like 5 times to remove overtraining detector

and he always tell me "just dont enable it"

😭

[latent cypress]

🤣

[edgy tangle]

lol

"If it doesn't work, just dont use it"

[analog obsidian]

he knows it's bad
i know it's bad
we know it's bad

yet still there

anyways that thing stops the training when the graph fluctuates a bit
which is normal during training

[latent cypress]

https://tenor.com/view/tae-hyung-bts-funny-funny-face-ohyeah-gif-12348224

[analog obsidian]

reaching true overtraining takes a fk ton of time guys

around 1.5k epochs

[latent cypress]

but if i maxed out the smoothing, the graph seems going down

should i always maxed out the smoothing?

[analog obsidian]

use 0.6

avg loss50 don't need too much smoothing

[edgy tangle]

1 just mixes and makes it look like it is going down

[latent cypress]

[edgy tangle]

looks overtrained

I think

[analog obsidian]

it's nearing convergence

it stopped going down at 20k

[latent cypress]

when it start overtraining?

[analog obsidian]

at 2k epochs

now the model is doing small adjustments

[twilit forge]

are u locally training?

[latent cypress]

yes

[twilit forge]

did u do the preprocess dataset in a zip file or just a folder with wav files?

[golden walrus]

wait, if i have a data that 39 mins long, it will get overtrained around 200 right ?

[analog obsidian]

guuuys

i told ya

[latent cypress]

2k? isn't that too early?

[analog obsidian]

true overtraining happens at 2k epochs

[analog obsidian]

2000 epochs

[golden walrus]

fork, i will train it to 200 and see then

[analog obsidian]

what are you seeing in your graph, is your model being close to converge

[golden walrus]

thank you

[analog obsidian]

meaning it's stopping having meaningful improvements

when it reaches the point of flattening you can keep training for a bit (1 hour or 30 mins) and then you can start hearing the epochs
just choose one that doesn't sound robotic

[twilit forge]

my preprocess taking ages

[latent cypress]

can you mark the sweet spot on my image?

[edgy tangle]

at 20k steps it started to overtrain ||(i think)||

@analog obsidian Im right?

[latent cypress]

[analog obsidian]

stopped going down after that

[edgy tangle]

yay im right

[twilit forge]

how long it supposed to take when i preprocess my dataset on applio locally

[edgy tangle]

how long is the dataset?

[twilit forge]

6 mins 30

[latent cypress]

what's your hardware?

[twilit forge]

im still on gtx

[latent cypress]

ah okay

[twilit forge]

1660

[edgy tangle]

It shouldn't take more than 5 minutes

[latent cypress]

is that enough to train on 1660?

[edgy tangle]

uhh

depends on the vram

[twilit forge]

im on 32gb ram

[edgy tangle]

not ram, vram

[latent cypress]

it's 6gb right?

[twilit forge]

yeahj

it might be 8 lemme see rq

[crude flame]

idk if this will help you but if you draw straight lines it can kinda help you spot where it starts to overtrain

[twilit forge]

yeah 6gb

[edgy tangle]

6gb it's not recommended, but should work fine with checkpointing

and for feature extraction shouldn't take too much time

[latent cypress]

owh nice advice, lyery was right, it's on 20k

[twilit forge]

cus ive been waiting for ages to preprocess the dataset

its been 15 mins

[analog obsidian]

usually the graph stays in a flat line and then goes up

[edgy tangle]

That is not normal

for 6.5 minutes dataset

[twilit forge]

thats what im thinking

[edgy tangle]

How many cpu cores did you set?

[twilit forge]

6

[edgy tangle]

set it to 1

[analog obsidian]

maybe you've got out of vram

[twilit forge]

how i check

[edgy tangle]

It starts using shared vram

[twilit forge]

ohhh

[edgy tangle]

And becomes too slow

[twilit forge]

do i need to restart the preprocess when setting it to 1

[analog obsidian]

check your gpu vram usage, if it's almost full then its using your ram

[twilit forge]

or just let it keep going

[latent cypress]

yeah could be out of vram like me who tried batch size 8 using checkpointing and nothing happens for 7 mins

[analog obsidian]

restart the whole thing and use only 1 cpu core

[twilit forge]

bet

and do i need to put vocals in zip file

[analog obsidian]

nop

[edgy tangle]

local?

[analog obsidian]

put them in a folder

[twilit forge]

yeah ok

[latent cypress]

if you on local, no

[edgy tangle]

If it is just a single file, use dataset creator hehe

[analog obsidian]

you can also put the folder in assets > datasets

[edgy tangle]

yeah

[twilit forge]

i did

[blazing solar]

Guys how to use kaggle?? Im newbie

[twilit forge]

i already sectioned out the dataset on audacity

do i keep audio cutting and effects on

[edgy tangle]

uhh

[analog obsidian]

merge all audio files into a big one

and use this in audacity

after that use simple mode slicing

with these settings

[twilit forge]

mine dont got this

[analog obsidian]

ah you're using the compiled version

[twilit forge]

yeah

[analog obsidian]

download main branch instead

[twilit forge]

the other version wasnt working

[edgy tangle]

Download main brach file

[twilit forge]

can u send the link?

[edgy tangle]

https://github.com/IAHispano/Applio

[twilit forge]

goat

[analog obsidian]

https://github.com/IAHispano/Applio

[edgy tangle]

Click code, then Download ZIP

[analog obsidian]

download the zip, place it in a non onedrive location

[latent cypress]

is this a must? because i always cut my audio in fl studio before putting it to rvc

[analog obsidian]

after that, use run-install.bat

don't close the cmd until everything is installed

when it's done, you can then use run-applio.bat

[analog obsidian]

yes

slicing does impact training a lot

[latent cypress]

so that's the problem why i can't get lower value

[twilit forge]

okay ill let u know when its done

[analog obsidian]

mainline auto slicer tries to slice the dataset into 3s chunks too but most of the time fails

[crude flame]

3 second slicing ❤️

[twilit forge]

are u saying on local traning its better to keep the file as 6 mins instead of slicing it?

[analog obsidian]

no

look

first have one audio file

if you have multiple, merge them into one

[latent cypress]

mine

[analog obsidian]

after everything is merged, use this

[edgy tangle]

No, just put the single file on applio, and then slice it with simple slicing

[analog obsidian]

then use simple slicing with the default settings

[twilit forge]

ohh okay

[analog obsidian]

it's very important that the silence is truncated with the settings i sent

[edgy tangle]

If you keep the silence, it will take more time at training, trying to learn about the silence (not good for gpu's health )

[latent cypress]

is there a way to automatically merge this audio? it's a lot

[analog obsidian]

merge them in audacity

[edgy tangle]

Don't you have the og file?

[latent cypress]

I ALREADY DELETED IT

[edgy tangle]

[latent cypress]

all i have is sliced audio

[analog obsidian]

just load everything into audacity

[latent cypress]

https://tenor.com/view/lewissaunderson38-lewissaunderson-cry-crying-cry-meme-gif-11499408157122608476

[latent cypress]

ah okay

[analog obsidian]

select all tracks and use this

align end to end

export it as a single file

then open it, run truncate silence

and train that

[latent cypress]

pretty sure my pc would straight-up roast me if it could talk, all this non-stop training

[twilit forge]

what should i set the silent training files too?

[edgy tangle]

Leave it at default

[balmy blaze]

Pls help me with kaggle

[twilit forge]

1000 epochs for my file should be good?

[edgy tangle]

6.5 minutes dataset right?

[twilit forge]

yeah

[edgy tangle]

set it at 300-350

[blazing solar]

What should i do after this?

[twilit forge]

and do i need to turn checkpoint on

[edgy tangle]

yes

4 batch size

[edgy tangle]

install everything

[twilit forge]

cache dataset gpu?

off

[blazing solar]

Where?

[latent cypress]

tf i've just added to my dataset😭

[edgy tangle]

https://docs.aihub.gg/rvc/cloud/applio-kaggle/

[latent cypress]

can rvc read that?

[edgy tangle]

just follow the guide

[edgy tangle]

i dont know, but I recommend you to put it at .wav

[twilit forge]

@edgy tangle

[edgy tangle]

off

try to put epoch saving at <10

[blazing solar]

Im stuck at 2 B

[edgy tangle]

like 5 or 1, so you can find the best epoch precisely

[twilit forge]

alright bet

[analog obsidian]

you can only find the best epoch saving every 1

[edgy tangle]

just scroll down the notebook sidebar

[analog obsidian]

every 5 or more it stops being precise and you're just using random epochs

[edgy tangle]

[burnt saffron]

curious if anyone knows a fix or if its just an issue with the dataset itself?

[analog obsidian]

increase batch size

[edgy tangle]

batch size >8

[burnt saffron]

alright thanks

[edgy tangle]

but for 10-15 minutes dataset its not "recommended"

[analog obsidian]

a problem with increasing the batch size is the lack of generalization in the model

so keep in mind that

[burnt saffron]

is there a specific batch size youd suggest?

[edgy tangle]

4

[analog obsidian]

try 8

[edgy tangle]

but if you want dataset fidelity, then 8

[analog obsidian]

noobies recommend 4 for everything below 30 mins

[edgy tangle]

at the cost of worse generalization

[blazing solar]

After that?

[analog obsidian]

this can be annoying

[edgy tangle]

click session options

[burnt saffron]

okay thanks

[analog obsidian]

also use a index value of 0.5

the voice will sound even closer to the dataset

[blazing solar]

Then run?

[edgy tangle]

then set this settings

[twilit forge]

is there any other setting i should play around with when traning the dataset?

[blazing solar]

Done

[twilit forge]

im saving each epoch as 1

[analog obsidian]

use batch size 4 for your small dataset

[edgy tangle]

do you already have the ngrok token?

[crude flame]

😭 i dont think adding 33 more minutes to the set helped my model

still need to mess with settings to see if i can get a better model

[blazing solar]

No

[twilit forge]

i was having the same problem as @burnt saffron cus im using the same vocals as him

[analog obsidian]

uhh the dataset has room reverb? i know that messes up things

[edgy tangle]

create a ngrok account

[analog obsidian]

use batch size 8 then

[twilit forge]

aight bet

[analog obsidian]

but your model will have bad generalization

[crude flame]

no, it doesnt have room reverb

[analog obsidian]

have you tried using a bigger batch size?

[twilit forge]

what u mean

[edgy tangle]

so use batch size 8, but you'll get worse generalization if your dataset is too short

[crude flame]

not yet 😔

ill try 12

[analog obsidian]

the model will randomly glitch

when trying to inference things too different from the set

[edgy tangle]

like it will break at trying to create its own voice

[twilit forge]

oh i get u

[analog obsidian]

good luck 🥹

[twilit forge]

would batch size 6 maybe help work around the problem'

[analog obsidian]

lower batch sizes are able to do more stuff but usually they start to sound different

[edgy tangle]

and sound horrible at inference when you use a voice too different from your dataset

[analog obsidian]

yep

[twilit forge]

okay thank u

[latent cypress]

@analog obsidian is this the glitch you're talking about?

[edgy tangle]

It just sounds with robotic sibilances

[analog obsidian]

glitching as the voice literally breaking

[latent cypress]

oh i see

[edgy tangle]

yeah

the model will noticeable die

[analog obsidian]

lower batch sizes generalize better but sometimes the voice tend to sound a bit different

[blazing solar]

Done

[edgy tangle]

not just sounding robotic

[edgy tangle]

then copy your token

and set it there

[peak shale]

my bum itchy

[twilit forge]

this is the last model it made the voice higher than the actual dataset voice

[analog obsidian]

bigger batch sizes are more accurate to the dataset, but the model loses a lot of generalization

[blazing solar]

After this?

[latent cypress]

tf am i supposed to do?😭

[edgy tangle]

?

[peak shale]

get ai to scratch it for me

[edgy tangle]

rvc is not just "perfect"

[latent cypress]

https://cdn.discordapp.com/attachments/1244363710986715247/1270699138581008405/ag.gif

[blazing solar]

Is this the token?

[edgy tangle]

no

[crude flame]

smh it needs to be perfect 😠

[edgy tangle]

PLZ

[peak shale]

boy i’m not reading all ts

[blazing solar]

I donno where is the toKen

[crude flame]

ive spent like a week just training one voice and i cant get it good enough 😡

[edgy tangle]

in the sidebar there is a section called my token or somewhat like it

[peak shale]

i trained my dog to sit down so

[twilit forge]

yeah i get u

[edgy tangle]

probably sounds like the pretrain

[analog obsidian]

@latent cypress @edgy tangle @twilit forge bs16 = batch size 16

bs4 = batch size 4

[blazing solar]

My authtoken?

[edgy tangle]

yes

[peak shale]

boy ts big ahh sa

[analog obsidian]

this is generalization

[edgy tangle]

Yeah, it is worse on batch size 16, because its too different

[latent cypress]

why sound so broken on 16? the dataset is not too long?

[twilit forge]

ohhhh

[analog obsidian]

batch size 16 was too high for that set

[edgy tangle]

the dataset doesn't get that note

[peak shale]

yo how do i vocal remove vocals

[latent cypress]

ah

[twilit forge]

too high for dataset

[peak shale]

isaak this is a real problem

[twilit forge]

mvsep

[latent cypress]

uvr

[peak shale]

can i send you a link and you do it

[edgy tangle]

with bs4 it generalizes better and try to fake the voice to reach that note

[blazing solar]

Just copy the "your authtoken" ?

[edgy tangle]

in the page there is a token

[twilit forge]

i got u 😈

[edgy tangle]

just copy it and paste it on kaggle

[burnt saffron]

blud

[burnt saffron]

NO

[edgy tangle]

what

[blazing solar]

I dont see this only command line & config file

[peak shale]

i wanted the nair video but it didn’t work

[edgy tangle]

bro

[blazing solar]

💀

[latent cypress]

i thought it was nair video

[peak shale]

it could’ve been

stupid discord

[edgy tangle]

the picture i sent to you its the kaggle page

[crude flame]

oh i forgot to change the lr back to 1e-4

time to train this again but correctly

[edgy tangle]

just copy authtoken and then paste it on kaggle

[peak shale]

guys who is this?

[lunar perch]

i try using f5-tts with german language.
download "marduk-ra/F5-TTS-German" from huggingface.
set the path to the files in the f5-tts webUI.
but the output audio sound like it is spoken reverse or maybe chinese.
any idea what i did wrong ? 🙏

[blazing solar]

wtf💀💀💀

[burnt saffron]

thats me!

[peak shale]

hey

i know you

[lunar perch]

i see, this discord is more for kids 😂
my fault 🤦‍♂️

[peak shale]

boy

[burnt saffron]

no way this mf just left the server wtf

[peak shale]

we made him leave boys 😎🥶

[blazing solar]

Where

[blazing solar]

Where to paste

[edgy tangle]

read it

[blazing solar]

What is wrong with this?

[edgy tangle]

you clicked start?

You need to install first

or it will not work

[blazing solar]

Install what