I have to use a vocal that someone overloaded their mic pre on and they don't want that. I'm pretty sure it doesn't exist, but I'd be happy to be corrected: Is there an AI thing out there that can remove distortion from an audio file? I've looked but haven't been able to find anything. It's not clipping distortion (I tried Izotope De-Clipper, no dice). An overloaded mic pre can be nice in other circumstances, but not for this clean female ballad.

Not yet.

Not in 2025 march. Saturation is non linear and impossible to predict or inverse.

So AI can do stem separation but removing saturation is impossible? I kinda doubt that.

Overloading destroys information. Where do we find the audio information which was never recorded?

By training the AI model on clipped vs unclipped audio?

10 years ago i would have agreed with you but not anymore.
Have you seen AI graphics upscaling? Same principle.

I've heard someone say that Spectralayers de-distortion module was pretty good with this. I haven't tried it so I'm not sure if it actually fixes distortion that isn't literally clipping, but might be worth a try.

Not yet.

You have to begin by coding a container set of code that deals with that analysis approach.

The few hundred guys/projects at the moment are pretty firmly focused on demix, clone, etc....which share basic code starter points to create specified focused model code.

A few hundred/thousand guys in the project communities will get un-distort routines explored/shared....but I haven't noticed anything workable yet. I give it over the next 4-5 years.

Demix/clone code still has a loooooonnnngggg way to go and that seems to be the area of focus that jazzes coding guys atm.

I will be open to you teaching me how to remove saturation with a plug in mate. I will gladly learn.
We are being able to separate, but not to recovered what hasnt been recorded yet.
Who knows.
I havent had any luck declipping some distorted recordings i had.

Didn't they remove the distortion from Lennon's vocal on "She's So Heavy" digitally for the 2009 remaster? (I think that was the song.) Is that the kind of thing you're talking about?

It seems theoretically possible for an algorithm to "guess" what the waveform would look like without distortion.

I don't even know what that means.
So you agree with me that it probably isn't impossible, it just hasn't been done yet? Which was what i was protesting against (that it couldn't be done).
I never said a plugin exists today that can do it. I was objecting to the notion that one would never exist.

AI graphics upscaling works by taking a say 4k image, downscaling it to say 1080p and train on that data. I don't see why you couldn't do the exact same with unclipped and clipped audio? The exact same principle. I obviously haven't tried it but i don't see why that wouldn't work.

And i'd argue that removing drums and vocals is a way harder task than removing distortion.

10 years ago i was convinved removing reverb from a signal was impossible but now i know it isn't.

https://borisfx.com/products/clipremover/

I have no idea if it's any good or if it's a scam even.

For the nerds:
https://joimort.github.io/distortionremoval/

Initially, as an example of approach...it'll be more this following way.......

1. Listen to the Greg Lake vocal on Crimson's 21st Century Schizoid Man....a vocal that was purposely recorded distorted.

2. Now, listen to the isolated vocal of Greg on the studio version of elp From the Beginning or Lucky Man.

3. Code will....in initial forays of this.....analyze Greg's non-distorted voice and re-build his clean sonic fingerprint into Schizoid.

Not quite the same as cleaning up an entire mono mix of Kingsmen "Louie Louie"....but ...in terms of state of the art of circa 2030-ish code ...that's the first direction code will do....imo.

While I am disappointed with everything they have so far, I agree it will be possible someday. This is not from any deep knowledge of how they work, but just from my experience of a lifetime of watching "impossible" things eventually becoming part of my everyday workflow!

But the thing about something like this is that it will always be guessing. Say you zoom in on a car from outer space and that car's license plate is one pixel in your camera. Well an AI can "zoom in" further and 'create' a reasonable-looking license plate, but it won't be the actual license plate. It won't be the actual numbers of that car. From one pixel, it would only be guessing at the state. The AI is creating new information and passing it off as 'enhancement' or 'repair'. It is not really being "extracted" - it is being "extrapolated".

Philosophically problematic and not admissible in court, but it could be just fine for "art", in the sense that the singer might not care if that's what he actually sounded like on that day, as long as it sounds less distorted.

Even though you say your file doesn't have clipping distortion, this might be worth a try, as it appears to be different from competitors:

https://myaudiofactory.com

Well damn, that's exactly what I need. Thanks for turning me on to it!

Yes it can be done

What about cowbell removal? Can that be done? I have a song and the producer swore to my band that we were going to want the cowbell, but as it turns out, the cowbell in the song is very distracting. I want to either turn it way, way down in the mix, or maybe even remove it altogether.

Darn that Bruce Dickinson. Should have never listened to him.

Just downloaded the free trial of this plugin. Slapped it on a female vocal track that was badly distorted, sounding like her mic was connected to a fuzz pedal.

Sorry, the plugin didn't remove the distortion. Because the vocal track was distorted, but it was not clipped.

There currently is NO plugin that "un-distorts" distorted audio.

Just talking out of my rear-end here, but if it's possible, I think it would have to be something more along the lines of Melodyne, where it operates over a specific range. It would have to sample another range that isn't distorted to get the non-distorted tone, then superimpose that over the distorted area, and make all sorts of little adjustments to make it sound natural. Otherwise, without having that more pristine range to reference, I can't see how it would know would know what true the reference tone really is. I don't imagine you could just cut out out the distortion like noise artifacts when the source is fundamentally changed. AI might be able to rework the distorted bits by inferring from other , non-distorted clip ranges though.

But someone smarter than me at this stuff may know better...

I agree. It seems to me that in distortion and clipping, information is genuinely lost. That information will probably have to be 'fabricated' - as opposed to 'reconstructed'.

hey that gives me an idea! .

You could invite the original singer to come back into the studio, set up the same mic, give them headphones, play the track, and ask them to give you clean samples of their voice to teach the AI what the target sound should be!

Accentize dxRevive does AI speech resynthesis that can work wonders. I don’t think it currently is able to correct distortion, but it doesn’t need the user to supply a reference tone—its target of what well-recorded speech should sound like that has been built into its AI model. It can make turds into a vocal that sounds like it was perfectly recorded.

“Our algorithms have undergone training using an extensive array of languages. They have been exposed to a diverse range of linguistic nuances and accents from speakers around the world through our custom-recorded speech database. Regardless of the input speech type, dxRevive excels at accurately restoring the intricacies of various pronunciations.”

Un-distortion might not be possible yet, but AI will get there eventually.

Greetings.

I saw this at a trade show in January.

They presented an example of clipped speech that was restored to not quite pristine, but decent condition. However, it was a trade show and it was their presentation, so I cannot vouch for the legitimacy of the effect.

Best

InSomnius

Mathematically that would be sorta true but remember two things: we're not dealing with completely random signals. If you look at a clipped waveform in an editor you can clearly see it, right? And so can the AI if you train it right.

And we're not after exact reconstruction (well that would be ideal though) more like the original than what we have is good enough.
I'd say that's more filosofy than engineering.

Sure because a clipped waveform is only missing information (flattened) at the peak. Using the pencil tool, I have drawn in the missing peaks myself on occasion. I am just doing it "by eye", and of course the little 'zigs' and 'zags' of the original peak are gone forever, but at least something is there rather than a flat line.

People are reporting decent results correcting clipping, but not correcting distortion which apparently is a different animal. Probably because the information is missing throughout the waveform, and not just at the peaks.

I can't complain too much, because I am from an era when many of the fixes we take for granted today were not even considered possible.

I have used this online AI service to remove drum bleed from a vocal track with pretty good results, so I would think it's possible.

https://www.lalal.ai/

Frankly, I was very surprised by how effective it was. I had like four songs done, and on one of them I had a little artifact on one phrase, but the rest was smooth enough. I went in not expecting much, and came out impressed.


edit:
It could also have been this one: https://auphonic.com/

I have receipts for both, but don't remember which one I ended up using.

On the other hand, a distorted waveform still contains all the information so that should be possible to completely reverse as opposed to a clipped one.

Here's some examples of undistorting AI from the page i linked above.
Original:
https://drive.google.com/file/d/1xH4...ew?usp=sharing
Distorted:
https://drive.google.com/file/d/1M5z...ew?usp=sharing
AI Undistorted:
https://drive.google.com/file/d/14pc...ew?usp=sharing

I am not so sure that it does.

Not bad.

It got most of the distortion out. It's still a bit dirtier than the original, but the amount of residual distortion on the AI track is more than acceptable - on an electric guitar.

A better challenge would be a lead vocal. Or some classical piano.