Just starting out with this stuff and want to minimize time wasted.

Q: Is it good enough to have the voice you want to clone just speaking - not singing - for the purposes of cloning and then having their cloned voice singing-voice on your tracks?

Got a great interview of a popular artist - well-mic'd and clear (better than extracting a reverbered vocal singing on a track using UVR, in other words.). Will this 'interview voice' translate well into a cloned 'singing voice'? (I have permission from the deceased artist's brother by the way, so cool on that front!)

I'll be experimenting a bit with this and let you know soon my results, but just curious if anyone else has gone this route - using a spoken-words voice for a singing clone model?

BTW: Here's a great vid of how to create your own voice models natively (off-line) using software you download to your PC:

ps: using the above software/python setup, don't be an idiot like me: make sure your folders (including your subdirectories) have NO SPACES or the first part of the cloning will not work. RTX is better/faster than GTX, but as long as you have GPU, you're good to go.

 

Last edited: Dec 26, 2023

It's an art to make a talking voice sing. Not impossible, and Melodysheep did exactly this:

 

But damn close. If @OP, you want to minimize wasted time, you should realize that this will be a rabbit hole. You could figure out a way to introduce some spoken-word into your project to use the vocal parts you have without reinventing the wheel.

 

I'm not sure if I understood your question correctly, but is it, 'Is it possible to make a spoken voice sing, that is, with a dialogue?' If that's the question, I did it with a friend who likes Canserbero. Although it's not entirely 'singing,' there were parts where he added small melodies, and it turned out well. I used Colab for this, meaning I trained my friend's voice only with WhatsApp audios. Then, I extracted Canserbero's voice from the song and replaced it with my friend's voice in the song, and it turned out quite well, especially considering it was a WhatsApp voice.

 

Yeah, NOT like this no. They extract and auto-tune the voice you WANT to manipulate that's already present. I'm talking about taking THAT voice (say, Robin Williams) and converting it to Humphrey Bogart's voice in the same role. But yeah, Melody Sheep is great!

 

Yep. This is what I be talking about lad. Currently baking a deceased artist's voice model in the GPU oven. Will let you know of the results.
Thanks!

 

Well, here I leave you a repository with various Colab links. You have the free version or anti-ban in two steps: first, download the repository in Colab A, and in Colab B, train and save. It's intuitive. You also have other options if you have a paid Colab account. There are links and videos for making inferences; it's quite comprehensive. I've trained several things here. Give it a try; hopefully, you can achieve what you desire.


Edit: Oh, and I want to clarify the 'anti-ban' thing. When you are a free user and download the repository or if there's any code that suggests to Colab that you are trying to train a voice, everything restarts. Colab wants you to pay for the work, in other words. But it's not that they ban your account. It has happened to me several times trying various Colabs, and it detected the code, but it only 'bans' the project you were working on, so you can rest assured with the term 'anti-ban

https://github.com/ardha27/AI-Song-Cover-RVC

 

This guy produced an album with Disney characters singing: I love it!

 

Beautiful, but those characters sing in the cartoons.