Could AI improve bad recordings?

[soundwave76]

We have seen lately impressive feats what AI (machine learning, neural nets etc) can do to images, from generating so called 'deep fakes' to improving bad quality images. So I was wondering could this be done to bad recordings with for example low dynamic range?

Thoughts? Definetly not an easy task, but impossible?

 

[DKT88]

Not impossible: https://technikjournal.de/2018/07/25/artificial-intelligence-in-music-production/

 

[soundwave76]

Did some searching and found these:

"MACHINE LEARNING APPROACHES TO DYNAMIC RANGE COMPRESSION IN DIGITAL MUSIC PRODUCTION" http://elias-mistler.de/downloads/ml_drc_final.pdf

Listening without ears: Artificial intelligence in audio mastering
https://journals.sagepub.com/doi/full/10.1177/2053951718808553

 

[DKT88]

Recording engineers are also artists (or some are) so the philosophical question is when will AI become artistic or accepted as such. There is no answer at this point.

 

[ScienceRaven]

I think it would be really easy to denoise it... also would be easy to compare 1920s/ 1950s audio to 2020s CDs and convert one to the other...

The problem is that very few scientists understand high definition audio processing To make libraries that everyone can use to do audio machine learning projects.. There is a wave to vector librand it probably doesn't even use high definition audio.

good audio is 44000x1024 Per second... I think you would only need 4 songs in Adversarial network so that the computer can convert the old to the new one... Whereas image learning projects currently use billions of images...

Analysing 4x 3min songs would be Equivalent to 80k images if we use 2048 freq bands...

It will be much easier to restore voice recordings and I don't think even Google has had the time to do that yet but when someone sets out to do it they should be able to do it very effectively.

I would recommend using wigner transform and not FFT

 

[USER]

Perhaps the most fully realized use of AI can be found in Peter Jackson's new Beatles documentary, Get Back.

They were able to "demix" mono rehearsal recordings in order to create the film soundtrack and also to listen in on conversations thought fully obfuscated by the band until now! The idea isn't new but his budget allowed them to create a better program for it.

He got a lot of flack for a previous documentary in which he colored WWI footage--lots of conference roundtables in film and documentary studies conferences--but here the context is very different.

 

[ScienceRaven]

Oh fantastic! I found info on progress of demixing here, theres a AI study of the beatles although its paywalled.
It would be awesome to know how they trained the neural network, academics are all probably running GAN2.
To advance audio ML, you can buy discs of acetate and vynil in very high audio quality, and then scratch them with wool, acids, sand and whatever so they become authentically old, and tell ML to convert old ones to new ones... Trick is devising authentic ageing. Also, to record high-quality audio through 1920s, 60s microphones and demagnetize them... Then ML community has high quality comparative sources to run ML from and the field can perhaps progress a lot.

 

[Frgirard]

Oh fantastic! I found info on progress of demixing here, theres a AI study of the beatles although its paywalled.
It would be awesome to know how they trained the neural network, academics are all probably running GAN2.
To advance audio ML, you can buy discs of acetate and vynil in very high audio quality, and then scratch them with wool, acids, sand and whatever so they become authentically old, and tell ML to convert old ones to new ones... Trick is devising authentic ageing. Also, to record high-quality audio through 1920s, 60s microphones and demagnetize them... Then ML community has high quality comparative sources to run ML from and the field can perhaps progress a lot.

it's a publicity for ML but which ? Martin Logan, Mark Levinson...

 

[voodooless]

it's a publicity for ML but which ? Martin Logan, Mark Levinson...

Machine Learning

 

[Frgirard]

Machine Learning

cool! you use the good expression.
when the thing will translate "i am a donkey" or "ape money" correctly we will call artificial intelligence.
and do not forget, this is human who create rules used by the things. if the human don't know how to treat sound failure, the thing can not do.

 

[voodooless]

if the human don't know how to treat sound failure, the thing can not do.

That's the fun part: that is not always true. If an AI learns to do things, we humans have no idea how it does it. It's basically a back box. Of course, we can guide the learning, and tell it how to look at a problem so it has an easier job, but in the end, you have no idea how it exactly works.

 

[JayGilb]

Recording engineers are also artists (or some are) so the philosophical question is when will AI become artistic or accepted as such. There is no answer at this point.

I personally don't consider recording engineers as actual engineers. I believe the term comes from the days when they were called on to repair and design studio equipment.
Larger studios have technicians to repair and maintain equipment, but are not called engineers.

 

[Frgirard]

That's the fun part: that is not always true. If an AI learns to do things, we humans have no idea how it does it. It's basically a back box. Of course, we can guide the learning, and tell it how to look at a problem so it has an easier job, but in the end, you have no idea how it exactly works.

i strongly disagree.

the black box is a fantasy.

 

[Frgirard]

I personally don't consider recording engineers as actual engineers. I believe the term comes from the days when they were called on to repair and design studio equipment.
Larger studios have technicians to repair and maintain equipment, but are not called engineers.

they are not but they do like if they was.

 

[KikoKentaurus]

Just reminding that we had a guy who tried himself to create this type of AI:

Where is the AI audio upscaler?

@Sam Ash it's up to extensive use of tool chain (example; MPC classic + shredder effects, MadVR... System) same as with music and again source dependant (no amount of post processing will fix a bad one). Unfortunately we don't have cross compatible ones (video will be pronned back to OS default...

www.audiosciencereview.com

 

[voodooless]

i strongly disagree.

the black box is a fantasy.

Great for you. Now some actual arguments please..?

 

[Frgirard]

Great for you. Now some actual arguments please..?

autorrity argument.

 

[voodooless]

autorrity argument.

So, no argument

 

[Frgirard]

So, no argument

you are a voyeur and will not talk about my professional life.

 

[voodooless]

you are a voyeur and will not talk about my professional life.

Eh.. okay?