Hello audio lovers...
I'd like to ask if anyone knows exactly what types of software, tools and voices various AI sites use to create their songs.

We've got a bet going among friends: Some maintain they exclusively use GM instruments (like Xpand or similar Roland/Yamaha) and voices trained with software like Replay and similar.
Others, however, say no, meaning they use most advanced instruments like Kontakt and specific plugs AU-VSTs/Synthesizer V.
Does anyone know something more precise and confirmed about this? Perhaps there's an official page where one could learn?

I’m waiting for your opinions to win the bet...

 

Last edited: Jan 11, 2026 at 10:31 PM

AI don't create songs like humans do. They don't use virtual instruments.
Engines take real songs, analyse them and "concatenate" audio elements on demand.

Here is a beginning :
https://towardsdatascience.com/audio-diffusion-generative-musics-secret-sauce-f625d0aca800/

Of course, it may sound like GM or Kontakt instruments, because real songs use GM, Kontakt ...
But AI don't generate MIDI notes to feed a virtual instrument.

 

Soundful has a sound engine as far as i know, with their own synths and sound library. Its more of an algorithmic generator, they use machine learning models to generate midi then use that to generate a song. There might be using direct ai models for some of the options like the vocal chops generator.

 

I don't know, I'm not sure...
This is because, listening to many songs created by AI, the instruments it uses seem to be very, very similar to each other.
In fact, out of 30 test songs, AI always uses the same 2/3 guitars, the same 2/3 synths, 2/3 drum sets, 2/3 pianos... The same ones and the same ones again... and this is strange.
Of course, it depends on how you ask AI, and maybe a slight change of instruments would be possible, but this is something that makes you think it's taking sounds from its own GM banks or something similar and not sounds from the file AI it's copying from.

Even assuming the AI uses a .wav/.mp3 file of famous songs to create its own modified version, this doesn't explain why it then always uses the same sounds to reproduce it... and this also lets you think that it takes the sounds from GM banks or similar. But which ones?
Those who train the AI models and their connection synapses should well know how it does it.

By now there are dozens and dozens of AIs, so some information must have got out of the bag... Doesn't anyone have more precise infos?

 

Last edited: Jan 13, 2026 at 10:51 AM

@ElMoreno, why torture yourself with AI? Use your time more efficiently, do without AI, and make your own songs instead!

 

I haven't quite figured out exactly how it works yet, but I think the workflow must be very similar to this...

 

Have you ever won a bet and collected the money? LOL

 

So, have you ever bet someone to create an AI song?

I never bet; I save the money I would have spent on bets, and with that money I buy something.

 

No, the bet was about whether AI uses GM banks or not.

 

And did you win or lose? What did you bet on?

 

Well, AI is trained on existing songs, so it’s going to learn the sounds of the synths which were used in those songs, Juno, Serum, whatever. And then use those synth sounds. Basically.

 

No one has won yet because we don't have official proof to show who is right.
I bet that AI uses a sort of GM banks to create its songs.

 

Yes, this is clear and I can understand it...
The strange thing is that AI - even after learning from songs made with Juno, Serum, etc. - still always uses those few instruments with the same presets to create its songs, plus 2 generic Room reverbs and 2 delays set to 1/4 (when it wants to do something special).
Sure, I understand that creativity is the domain of humans but... at least show that you've made an attempt but... nothing!

The only thing it does well (when it uses voices) is that it has learned the rules of Harmony perfectly. It always manages to place the notes correctly, and even when it rarely makes a mistake, it moves them up or down and immediately returns to the key or then follows the modulation.

 

Last edited: Jan 13, 2026 at 6:51 PM

Well, that's good then.

Just think about it: how do you steal something without getting caught? You generate what already exists and then obscure the traces with a converter. It's called "generated," and nobody really knows where the data comes from, except maybe the company boss.

There are plenty of freeware samples and enough presets and other sources where they can simply steal and disguise the origin by renaming or reworking the material. I've said before that if necessary, they'll pay artists to record something and then relinquish their rights, similar to studio musicians who used to get €4,000 per album but didn't own the rights to the songs.

In America, almost every employee has to sign a non-disclosure agreement!
And with Chinese AI? The party is always right, so nobody questions it.

 

That's exactly it. They steal and then hide the evidence by not using the sounds they've stolen but always using just the same few instruments... so it all seems homemade (this becomes much more obvious when you leave all the creation to AI).

 

Pay a guitarist to play all the notes and chords, using different effects, distorted and clean, plus humbuckers and single coils, and maybe even different amplifier types. That's how you train the AI; that's how I would do it. Do the same with singers, drums, and bass.

It's like with Google; they supposedly don't collect anything about you personally, but they still use your data.
We've all trained the Google search engine over the years.

 

in the learning process, whatever sources it starts with are going to end up averaged with other sources, so regardless of the number of starting components, you'd end up with only a few different outputs/varieties at the end.

 

I'm a guitarist, so I always play the guitar parts live. I'm mainly interested in AI for the vocals.

However - since I still haven't quite figured out how to do it - what I'm really interested in is that I'd like to train my voice (in the style of how Ace Studio does it), but no one has managed to explain to me clearly and precisely how I could achieve this.

Anyone? The offer is open...

 

AI took the R2R iLok cracks and cracked the UAD suite, 'Internal Only'... Either that or its Behringer.

 

Good summary:
https://towardsdatascience.com/audio-diffusion-generative-musics-secret-sauce-f625d0aca800/

https://arxiv.org/pdf/2306.05284

-Afaik modern Music AI models don't calculate notes, they just simulate/vomit the final Result..
(Relentless Doppleganger one of the first..)

However I guess elaborated music AI's like Suno could use a more multifactorial approach..

You could have one Agent specialised and trained in Percussions/Drums,
and train it with the exact sound kit you want it to replicate..

Then make it be part of the audio construction chain with other agents/instruments,
which will be handed the same "structure" to fill up themselves..

I imagine the models use their own "machine language" they have developed empirically (by training) which only they can understand..
The trained models are like a smol seed with an algorithm, where data is generated after a prompt, etc..