I understand your situation. The key issue here is the difference in computational power. Google Colab and systems with CUDA are designed for the advanced AI algorithms we use in 2023, offering far superior processing capabilities compared to older PCs and Intel Macs. Although these older systems are supported, they're not ideal for our current technology.

The performance limitations you're experiencing are due to the hardware constraints of your computer, not a flaw in our software. It's like comparing the speed of a 1990 Honda Accord with a 2023 Tesla - both cars, but with vastly different performance levels. Therefore, while I sympathize with your frustration, the slower performance is a result of the older technology in your computer, not our software.

 

ha ha another proof that devs are always bashing win98 users with so-called new tech and then blaming the pioneers who are still using win98 and vista. Full 16 bit audio has been available since the atari st so I don't know who kidding who here.

 

Ok bro. Now.....


Although I think the answer is 'no,' let me ask you this. Do you remember that in AZ, we talked about using this for more than just cloning a voice for fun and discussed cloning instruments? So far, all the Colabs I've tried and searched for in various projects from different people and teams can only clone monophonic instruments, such as saxophone, bass, a guitar solo, etc. Does your Colab that you posted here or any other have the capability to clone polyphonic instruments?

 

Feel free to make a version of the tech that runs on win98 etc. As you can see the community would love it!

 

You can utilize this technology to create various instruments. I've created Dubstep growls, Reese basses, flutes, and more. However, this technology is suitable for anything monophonic, such as a voice or a single instrument. To achieve a polyphonic effect, similar to a voice, you need to record multiple takes. This allows you to harmonize, create chords, or even produce a children's choir effect.

 

If not, well. But I was referring to polyphonic, like cloning an acoustic guitar. I'm not going to record the strum string by string; I've tried things like that, even using Melodyne to separate the notes, and the results are very poor. We'll have to wait for new advances in AI

 

In that case, I would suggest using guitar samples, virtual guitar plugins, samplers, or a real guitar. Some tools in the toolbox aren't meant to do everything.

 

Had a bit of fun and Trained up Lzzy Hale from halestorm

 

This is amazing!! I assume by your username you are a male? Did you use your vocals first? then +12 them to hit her range?

 

Hey everyone we have a public solution to training a model for Replay!! Use the link below. Please read the instructions carefully. You have to split, rename, and properly zip the audio to make the training easier. Follow the link below and read.

https://replicate.com/replicate/train-rvc-model/

Share the models here!!

 

Sorry @curtified , a bit OT. Ryck, wasn't you who made a thread about instrument cloning? From guitar to cello for instance
Can't find it. What AI was that?

 

Please...more simple instructions....
sorry but i dont understan where go the result.zip

thanks for your work

 

You can search on our website www.weights.gg for instruments like a Guitar (https://www.weights.gg/models/clm72zu7m1aizcctc8rkpo4m0) and use them in our app.. all in the orig post

 

Did you read the readme?

 

Uppps....................I can see it.
sorry.

All is clear now

 

No i did not use my vocals, and yes I am a male

The test was to see how good this process actually works with a trained set of vocals of Lzzy hale from Halestorm
I Iike top do things manually not really interested the all in automatic solution.

So lets break it down
The Lzzy Hale weights were trained as follows
Took a Acoustic track of lzzy hale from halestorm that she had sung demucd out her vocals
split the vocal into a dataset
Trained the weights from her vocals

Then I demuc the orginal powder finger track took the male singers vocals from the powerfinger track And ran the lzzy hale weights on the vocal i had trained with a pitch adjustment of +6 using the rmvpe_gpu weight set using replay.

The lzzy hale vocal track was split using instant data set software that automatically splits the track into 10 sec splits

The Setup was

Ultimate vocal remover 5.6.0 for the demuc process of both the orginal tracks to obtain both vocals

Free software
Ultimate vocal remover 5
https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v5.6/UVR_v5.6.0_setup.exe

Free software
Instant Dataset Maker to make the dataset of the Lzzy Hale vocal track

Download link : https://pixeldrain.com/u/N4wG1VCf


Free software
For training RVC1006Nvidia for training the dataset of the Lzzy Hale vocal track

https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/RVC1006Nvidia.7z

RVC1006Nvidia settings to train
Split Dataset.zip created with Instant Dataset Maker
40k,
Version 2
100 epochs
Rmvpe_gpu
T batch_size string -6

Weights training time was overall 30 minutes

Reconstructed the song in Reaper from the previous demucd instrumental backing tracks from the powerfinger track less orginal vocals with the newly appiled changed lzzy hale vocal track


--------------------------------------------------------------------------------------
so Now we move onto Ai for music as it seems to be taking off

I Also had the opportunity to sit with https://www.suno.ai yesterday and created a few snippets of what it can actually produce sino.ai is only in beta
imagine the future possibiltys


Iittle tip continue from your first clip generate the lyrics 2 verses at a time
at the end join all the clips togther as one


Destructive Tears - Rock wife

Destructive Tears - Cant Stop the Fire

Destructive Tears Higher N Higher

Destructive Tears - Rengade

 

Last edited: Dec 1, 2023

Yes!! you're using all the tools I mentioned in here!! Holy shit I love it! I do a very similar process for my AI songs!!

Again the one you made is amazing!!

 

Hello @Xupito, how are you? Yes, exactly, but remember that I only cloned a monophonic melody – it was the cello, a guitar with a slide, and in another thread, I think I posted one of saxophone and bass. The results are "good," but the AI still needs improvement. I believe the methodology should remain the same. What I use is a code from https://github.com/Retrieval-based-Voice-Conversion-WebUI. I use it in Colab, but since Colab is banning everything related to the word "Retrieval-based-Voice-Conversion," what I did was change the name of the folder, and in the Colab codes, I did the same. I named the folder "test," and the directory it points to in Colab is content/test/. Then, I save it in Google Drive to the folder, and every time I want to clone something from Drive, I transfer it to Colab. After that, I just install the dependencies in Colab, open Gradio, and I'm ready to start cloning or converting whatever I need. For me, Colab is much faster (depending on the PC). Besides, it serves as multitasking; you can download, for example, a package of voices in Colab, then with a short code, cut those voices, and then clone them right there – all in one


edit: Sorry, I think the link doesn't work. I assume this is the original one: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI. The thing is, there was an error in Gradio at that time, so I created an account, cloned the repository, and edited a few lines to incorporate the Gradio update. This is what I cloned, and it's the repository I currently use: https://github.com/RyckAS/Retrieval-based-Voice-Conversion-WebUI. The last time I tried to use another Colab (about a month ago), the YouTube videos and RVCs I found were banned due to their code, and it had to be done locally. Since I didn't want to go crazy just to clone a few choruses, I did what I told you. I'm not as involved in this now due to lack of time, so I'm a bit outdated. The original poster (OP) is doing a great job here. I hope that at some point, their application can reach more users.

 

Last edited by a moderator: Dec 1, 2023

With the utmost respect, this sounds very, very off. The least it resembles is a guitar; I see it as something more experimental, but not a guitar clone

 

Feel free to make your own guitar model and use it in replay. From above it appears that you know how to use the colab. Make a better model for our community than the one I suggested.

I also included a link to the entire search results on our community maybe there is a better guitar model that I'm not aware of. There are over 18K models and growing. Help us out if we don't have one. Or point one out that we not be aware of.