PG Music Home Forums General Discussion Help! Tech S.O.S (Off topic) Voice Changer

Does anyone have any experience with a plug in voice changer? I would like to sing on my tunes but I can't stand the sound of my voice.

Unfortunately, I have no actual experience, but I'm not sure if any of these fit the bill:

https://www.antarestech.com/products/auto-tune/pro

https://www.waves.com/plugins/waves-tune

https://www.reddit.com/r/ableton/comments/y8undk/best_vst_pitch_correction_software_for_natural/

https://audiosex.pro/threads/best-live-pitch-correction-vsts.70799/

Others will have more experience (I hope)

There's also this after-market tutorial - not maybe what you want, but interesting anyway:

Another option (which is a lot more work than a simple voice changer) is to use a synthetic voice. Most of them let you convert your voice into that program's internal representation (notes, phonemes, and pitch deviation), and then have the synthetic voice sing it.

There are a lot of downsides to that approach. Using a voice changer is basically loading the source voice, picking a target voice, and getting the results.

With vocal synthesis software, you need to:

* You have to buy/subscribe the software
* You have to learn the software (it's like a pitch correction software plus words)
* You have to find a voice you like
* You have to pay for that voice if you're not on a subscription
* You have to fix the import errors (pitch, phonemes)
* The performance may be very good, but still to some degree remain synthetic.

On the other hand, you have a lot more control over the voice. Voice conversion software leaves the original performance intact, including issues with timbre. Correcting these in pitch correction software only goes so far, and if you move the voice too far out of range, you'll get artifacts.

With vocal synthesis software, you can completely re-sculpt the performance, and even replace things like vibrato without the timbre going wonky. Changing the gender of the performer is also much more realistic.

You have quickly been introduced to two new very powerful tools. But before you go too far down the rabbit hole I would ask you to reconcider these options. I was in your exact position years ago.

What helped me to accept my voice was two things:
1) Vocal Lessons from a pro. There is a lot which can be done to improve how you sing which will improve how you sound.
2) Pre-vocal warm up exercises (see #1) followed by Post vocal processing such as Isotope's Nektar and Neutron.

I still have cringe moments with my vocals which require me to "redo" the performance and I have been accused of overusing the post processing (too much reverb...), but I do get some satisfaction for getting more of me in my music.
Think about it.

Nobody likes the sound of their own voice. Learn to live with it

All well and good, I suppose...

I am looking for an actual voice changer app that lets me use a voice of my own choosing. I'm tasked with trying to finish a project for a South American client. Unfortunately, the lead singer passed away before some of his tracks were recorded.

I don't have any working expertise here, but I do know that Dae Lims, the guy who made did an AI remake of SMiLE had ten different models of Brian Wilson.

In contrast, there were only two vocals models for Randy Travis' AI song - one used 12 vocal samples from Travis’ career, and the other used 42 samples from more recent songs.

I need something more basic such as to know which app and a how-to.

Samples are not the problem—I have them going back 35 years.

Sure.

Perhaps the most popular tool to use is RVC (Retrieval-Based Voice Conversion). The source code is available on GitHub, and there are a many tutorials on how to use it, as well as tools built to make it easy to install and run.

RVC can be installed locally on your computer, but it seems to have the most support on Windows, as opposed to Mac or Linux OSes. Once you've installed the RVC software package, it'll run locally as a web page. You can then do two things:

• Create a voice model by training RVC on audio; and
• Convert a source audio file using a voice model

It requires a lot of processing power to train a voice, so you're either going to need a computer with a powerful video card (to use the GPUs in the card), have a whole lot of time, run the training on a virtual machine, such as Google Collab, or hire someone to create the file for you.

On the other hand, converting an audio file using RVC can be done with a plain-old CPU. It'll even happen in real time if your computer's processor is fast enough.

There are a lot of tutorials out there, and even a number of forks of the various flavors of RVC. I haven't actually used RVC myself, but I'll supply links to what look like fairly reasonable tutorials. Cavaeat emptor, cave canem, etc.. You're a lawyer, so you love it when people write stuff in Latin, right?

I picked these particular tutorials because they appear to be targeting the most recent versions of RVC.

The simplest approach would be to use Google Collab. There's nothing to install on your computer, you don't need any GPUs, you can test it out free, and choose to pay when you need more processing done. What you will need is a Google account.

There is a web interface for RVC, and Google appears to be preventing Collabs with web interfaces from running in their "Free" tier. So there are a number of older RVC Collabs that no longer work in the "Free" tier.

Here's a video tutorial where someone's created a Google Collab for you. The video shows you how to run the Collab. It automatically installs RVC into the Collab. Once you've done that, you can load up training audio, train a voice model, save it to your Google Drive (so it doesn't go away when you close the Collab), and then apply the trained model to a source audio file. You can also retrain a model, and load a prior model:


Link to Google Collab: https://colab.research.google.com/drive/1hmKPNeeReO4NHzOktJFMI2plYKy5F7ZI?usp=sharing

Notice that he runs only 100 epochs of training to the model, as he's using a free Collab. It takes a lot of processing (and time) to get a high-quality voice. That's where paying Google for the Collab comes in handy.

I also noticed that he didn't split the training data up into a bunch of 10 second files, as most tutorials would have you do. It apparently speeds up the training time to use smaller chunks, but gives less coherent output. (This particular Collab's code seems to only accept one training file anyway).

This one Google Collab is all you need to create a high-quality RVC voice model, and apply that model to target audio. Easy peasy lemon squeezy.

If you want to run things locally, you'll need to install RVC on your machine. And you're likely to have the easiest time using a Windows PC instead of a Mac or Linux machine.

Here's a tutorial (among many) on how to do that. Again, if you don't have a supported GPU card, you might not be able to do any training (because some versions of the packages don't allow that), but you'll be able to perform conversions using a voice model locally:


Here's a tutorial on how to install a web-based RVC Voice Changer Client. It's aimed more for people who what to use RVC as a real-time voice changer, but it supports file input and output:


There's also free program called Replay that's uses RVC, but has a simpler UI. Here's a tutorial on how to install and use it:


Did that help?

Since I posted all those details, I figured I'd have a go a it myself.

• I used Spleeter on a set of early songs by America (Daisy Jane, God of the Sun, Ventura Highway, To Each His Own), and split them into vocals and backing track.
• In Audacity, I loaded the isolated vocals, and edited together Gerry Beckley's solo parts for 2:15 of medium/low quality vocals.
• I used the previously mentioned Google Collab to create a Gerry Beckley vocal model. It took about 25 minutes to run 200 epochs. A high-quality training would be around 1000 epochs, depending on the training data. Audio upload/download times to Google Collab are fairly slow.
• I saved the resulting model to my Google Drive (making sure the code didn't do anything malicious to my drive first!).
• I used Spleeter to isolate the vocals and backing track on America's "Tin Man", which is sung by Dewey Bunnell.
• I used the same Google Collab using the Gerry Beckley model to convert Dewey Bunnell's isolated lead vocals from "Tin Man". The result was as expected from using dirty stems and a low number of epochs. The voice was clearly Gerry Beckley, but with relatively low sound quality and some poorly converted phonemes.

Obviously I can't post the results.

I also downloaded Replay. Recall that at heart, Replay is just a wrapper around RVC:

• Since I don't have supported Nvidia video card, it could only run in CPU mode so it was slow.
• I needed to watch the video to figure out how to use it. It didn't help that part of the options can't be seen without scrolling the screen.
• I tried creating a voice model. It took Google Collab 25 minutes to create the low-quality model. In that same time, Replay only got three epochs completed, so I cancelled that.
• I downloaded the RVC voice model I'd created in Google Collab from Google Drive. It was a .gz file, so it required being decompressed several times using 7Zip. The .pth file contains the model, and is found in the \assets\weights folder.
• It took about 17 minutes to covert the audio file using the voice model.

So using the Google Collab is an easy way to try out RVC for free. The one I linked to is by no means the only or best one out there, but it is super easy and runs in the Free tier of Google Collab. And if you've got the paid version of Google Collab, you have access to the RVC GUI, where you can see how the training is doing, or restart training on an existing model.

In the other hand, if you've got a supported Nvidia card, it looks pretty easy to run RVC on your machine, especially if you use something pre-packaged like Replay.

I hope this helps!

Hey, David, this certainly points me in the right direction.

Not doing Windows anymore but my Mac Studio M2 Ultra's 76 Core GPU should be able to do the necessary crunching—if not, there's a 24 Core CPU.

Although I have a few Spleeter based apps plus Steinberg's SpectraLayers Pro 11, none of that should be necessary. I have a few GB of isolated vocal tracks from this and previous projects. In addition, another band member does a spot on impression of this guy's phrasing (usually, I'm saying, "cut that out—sing like yourself!") so all I need to do is to see if I can get a convincing match of his timbre.

I'll update with my progress. Thanks again!

Glad to hear it!

That's far more than my puny machine! For a moment I considered picking up a Nvidia card with CUDA so I could try running some of this technology locally, but it looks like I'd have to pay a minimum of $250. That's not a huge amount of money, but I've other toys I'd rather buy first.

Sorry, I was just describing what I'd done, it case anyone was curious. For zero dollars, it's possible to experiment and get interesting results.

With someone that's already able to do a convincing imitations, plenty of training data, and a high-powered machine, it sounds like you should be good to go.

This tool is VERY easy to use and I thought I got great results. https://create.musicfy.lol/?ref=joanne
I have not had much joy with SynthV

Here is a song I made using musicfy
[video:youtube]Stand by Me (capo 3) Cover Song by CooperSync play along with scrolling guitar chords and lyrics [/video]

That’s really impressive. What did you use for the harmonies in the chorus?

I used Realband to generate harmonies. Then I recorded these harmonies and then I applied musicfy to change the harmonies into other voices

This is really impressive Joanne. Enjoyed my listen.