RVC: Local Training Tutorial (How to make your own custom vocal a.i.)

my socials:

The shareable version of your path file will be in the “Weights” folder !!

hope this helps :)) this should be pretty easy and straight forward for ya.
try asking Chat-GPT for help! No joke

If you want to run So-Vits instead:

there should be a google collab tutorial linked there as well!

Chapters:

00:00 - What we will be covering
00:25 - RVC installation
00:40 - Explaining interface
1:11 - Isolating vocals with U.V.R.5
2:45 - De-echo or De-Reverb the vocals
3:00 - Preparing your Data-set
4:40 - Setting up Training
5:40 - Extracting pitch and features
6:10 - Explaining Training Settings
7:00 - What your model files will look like
7:40 - Have someone train a model for you

[LucidFirAI]

I dunno if you can add subtitles or notes to a video that already exists but...

KimVocal2 + VR DeReverb + VR DeEcho Aggressive + VR 5HP Karaoke work amazingly to isolate the main vocals.

I can't seem to use ensemble to do all 4 simultaneously to get one output (I can do 2 at a time, and I get multiple outputs rather than an output that used all 4 of the above effects)

A good tip I don't think you mentioned that is awesome is that you can load as many audio files as you want, so if you're taking a voice from an album you can load the entire album in and apply kim vocal 2 to all songs, then apply dereverb to all the kimvocaloutputs, etc.

(I'm using UVR5 separately, it doesn't seem to work in my RVC gradio)

[Mowgi]

Currently in the process of training my first model, thanks to you. Thanks P3tro <3

[BlueprintBro]

Fantastic tutorial, straight to the point with things. Now I'm just waiting for the training to end. 15 minutes of cleaned up voice, 500 epochs been 24+ hours 😅😅
Massive EDIT (If you are facing some issues like thinking the training is taking too long/stuck/suspended)

At first I thought the training just needed a long ass time. I left it to "work" for over 48 hours, only then did I check the command line, where it clearly stated that it was still at epoch 13, after this long, the program must have bugged out or something.
What I am doing now is saving a checkpoint every 5 epochs instead of 50. Since my first attempt did not even reach 50, it did not save, and I had to restart my model from scratch.

The point of this post is that YOU CAN RESUME training a model anytime you choose, as long as you have those checkpoints. (I did not know this, which made me hesitate to turn off my PC or check the command line before)
i.e. I want to train my model for 300 epochs. If it bugs out / stops for whatever reason at epoch 273. Stop the program, and restart it, using the same parameters you input when first training the model. It will load up the checkpoint at epoch 270, since I am saving a checkpoint every 5 epochs, so the last checkpoint would be at 270.

I noticed it would bug if I use too many threads, and a large batch file size.
For my specs 9 threads, and a batch size of 5 is optimal. 1 epoch would take around 4.5 minutes
Trained a model for 300 epochs using around 50 minutes of sliced up audio. It took around 23 hours.


Specs for those who are curious:
i7-8700
GTX 1080
Using 8 threads to process

[amcerngaming]

Thank you so much for this video! I spent days researching on YouTube on figuring out how to generate and train AI voices because I wanted to create my custom made voice, but this one had the most clear and straightforward instructions that finally got my project to work. I almost gave up until I saw this video ngl lol

[TheCommentVEVO]

Speed info:
3090 24gb - batch size 36
allocates 22gb - around 17 seconds per epoch

[martinfleiva]

I've been watching your tutorials, you're a god! Thanks a lot.

[Thyronix]

for those of you having trouble, he skipped over a bunch of important shit and got a some things wrong. First off, you can't load those pth files (he shows that it generates in that folder) in a voice changer client, you have to load the one in the "weights folder.." you'll notice it's actually named what you named the experiment... second, when you click "train model, " that doesn't generate the index files. You can either click "One-click training", or train it, then click "train feature index." Some more tips, I've had great quality using 700-1000 epochs and 10 minutes of clean audio separated into 10 second chunks. Don't use version 1 like he does, use version 2. Using more than 10 minutes of audio seems to slow the training by a lot and doesn't help with quality much. You can't use too many epochs though, it just won't get any better past a certain point. RTX 4070 12 gigs Vram took about an hour and a half to do 700 epochs.

[eto]

There is no .pth file after training. I see there are index files only, no ckpt either

[LosantoBeats]

Can you make a video on how to continue training the model. Covering things like adding more audio files to the dataset, and running more epochs on the model. Thanks!!

[Twiddle_things]

This guy chill asf damn. Thanks for the tutorial man!

[-nishizumisan-]

Только что подумал о том, что неплохо было бы, если петро запишет гайд по rvc, и тут же видео выходит. Магия.

[adallaire]

Hi p3tro! Great stuff works for me as explained. I'm blown away. Thank you for this. Will you do a tutorial about the 'ckpt Processing'? I'd be very interested to hear what this would sound like. Tx

[BeeBN]

now that google is cracking down on rvc this is a good alternative that they CANT take from us.

[ItsJustAdrean]

A note for everyone using this and getting an error, make sure there are NO SPACES in the names of your speaker, folder, etc! I don't know why it worked for p3tro and not for me, but this stumped me for a sec

[pujabanchu]

Even after completion of all processes, i cant see any .pth file for vocal i trained...i can see only index file..help me with that

[Mowgi]

So my first training took 12+ hours. Changed it to 24 threads (I have a 5900x) and the next one was done in 5 hours lol. Man, my fans were going crazy for 12 hours that first time 💀💀💀

[ThisIsntmyrealnameGoogle]

One of the best youtubers and one of the few I have notifications for thanks! Any good quality mics you recommend for hobbyists (non professionals)?

[MusikIndieNesia]

Hi P3tro... I was wondering if it is possible for us to train a model, then stop... then continue the next day... I was aiming for 1000 or more epochs. I guess my question is, does training have to be done in one go? Or can you stop and continue? Thanks P3tro!

[blackhand]

If you DO have a path file but no index file, don't worry. A fork of RVC called 'Replay' makes it so that you don't need an index file, just the path file.

[aschenixy]

For those having problems with step 2a having a blank preprocess note or processing data with blank output information... I think you have to make sure that the training folder doesn't have spaces or probably even symbols... Hope this will help as it did for me.