Am I staring extinction in the face?’: classical music and AI

[RichardT]

Getting back to the subject of the OP, one of the things that I think we lose sight of as people involved in the production of music, is that to the majority of the general public, music isn't that important. Even in the heyday of popular music most of the time it has been used as background noise. These people for the most part will be perfectly happy with AI generated music provided that it has a pleasant tune without being too intrusive. Also I would suggest that these people are not our audience and never will be unless we are suddenly to become very popular.

I completely agree! And I think AI will also make inroads into genres such as pop where acts have always been ‘manufactured’ for maximum appeal.

[Martin Walker]

Interesting news item on the radio this morning, apparently AI generated music has to be labelled and described as such, on internet downloads etc.

Interesting - I've seen plenty of 'AI-generated music SHOULD be labelled' thoughts, but it would certainly be good if this has been enshrouded in law.

[James Perrett]

Interesting news item on the radio this morning, apparently AI generated music has to be labelled and described as such, on internet downloads etc.

Interesting - I've seen plenty of 'AI-generated music SHOULD be labelled' thoughts, but it would certainly be good if this has been enshrouded in law.

I've been putting music up for digital distribution over the last few days and I was specifically asked about AI content for every track I put up. So, even if it isn't a current requirement, the distributors are taking it seriously.

[BJG145]

I’m very sceptical about this…it’s a very blurry distinction with a continuous scale between human or AI, and sounds unenforceable. (Just going on what I’ve read here.)

Surely the only way to manage it would be some mandatory watermark from the generators, which won’t happen.

[amanise]

Interesting news item on the radio this morning, apparently AI generated music has to be labelled and described as such, on internet downloads etc.

Interesting - I've seen plenty of 'AI-generated music SHOULD be labelled' thoughts, but it would certainly be good if this has been enshrouded in law.

I've been putting music up for digital distribution over the last few days and I was specifically asked about AI content for every track I put up. So, even if it isn't a current requirement, the distributors are taking it seriously.

I think that has a material effect on sync licensing in our case with the SOSFA LANDR setup. I have noticed that if you declare your track to be completely free of samples or AI assisted content, it gets synced straight to YouTube immediately. If you indicate that you have either royalty free or paid for samples, or AI assisted content - your track is labelled as unsuitable for sync licensing with Youtube. It then doesn't get distributed to YouTube Music. It's not something I understand - or particularly care about. YouTube streams pay 00.000 per stream anyway. But I use the LANDR AI Mastering tool, and EZD3 (which is MIDI driven drum sampling) - so it seems to me that I should tick both boxes. Again, it's not something I understand - but there it is happening for real.

[James Perrett]

But I use the LANDR AI Mastering tool, and EZD3 (which is MIDI driven drum sampling) - so it seems to me that I should tick both boxes. Again, it's not something I understand - but there it is happening for real.

They specifically talk about AI music creation - they're not talking about AI assisted tools where there is no creation involved.

Samples are similar - a huge percentage of synths use samples of some kind these days so they don't really count.

[RichardT]

Yes, by samples I think they mean bits from existing recordings, not the use of sample libraries.

The licensing for sample libraries lets you use those freely in commercial work. I don't think you need to declare those to your distributor.

[OneWorld]

I'm free on Wednesday, if that's any good to you?

Oh good! My bog could do with a scrub.

Dave.

Almost any debate descends into potty talk and our obsession with lavatorial humour would a robot plumber with the trots make a 'bolt' for the toilet?

[OneWorld]

I watched a programme on TV tonight, Channel 4, where the programme asked "Will AI take my Job?"

There were several individuals...Fashion Photographer, Lawyer, Doctor
Musician/composer

In each test there was a human expert and the 'equivalent' AI, then each of the works presented to an expert who would choose from the 2 examples, in a blind test, in each of the 4 scenarios, and say which they preferred, or were more convinced by.

The musician's brief was - write a short piece of film music 1 1/2 minutes

I wouldn't have said it was an easy brief, but that said, the AI did a reasonable job.

However, right from the first few seconds, he musician's attempt was clearly superior to the AI, which admittedly was good, but it didn't capture the visual message. The film director without hesitation went straight to the musician's offering which had far more dynamism.

I the AI's defence, it was written in about an hour with some AI subscription app costing £5.00 a month. The musician took all day and said he would usually charge about £500.00 for that day's work.

Over the whole programme, the human options won out, but only by a slim margin, the music example was the most convincing win. Seems we don't need hang up our plectrums just yet.

[adrian_k]

I saw that. I did feel the ‘composer’ using AI might have done a better job by making her increasingly complex prompt include timing information for hit points/mood changes, to match the film better. And in the end that’s why the musician won out, the director felt the music fitted the action better, not that the music was intrinsically better.

Coincidentally my son sent me this last night. It’s an AI generated cover of a System Of A Down (metal) song in a soul/jazz style. Pretty convincing.

https://youtu.be/nc99azuoFxo?si=p_iWkGqxCo5gEtdy

Here’s the original:

https://www.youtube.com/watch?v=iywaBOMvYLI

[BJG145]

Not bad for an algorithm. It's definitely alarming that the average listener probably wouldn't notice the difference now.

[RichardT]

It's brilliant. Scary. I prefer the vocals on the soul version to the original.

[OneWorld]

I saw that. I did feel the ‘composer’ using AI might have done a better job by making her increasingly complex prompt include timing information for hit points/mood changes, to match the film better. And in the end that’s why the musician won out, the director felt the music fitted the action better, not that the music was intrinsically better.

Coincidentally my son sent me this last night. It’s an AI generated cover of a System Of A Down (metal) song in a soul/jazz style. Pretty convincing.

https://youtu.be/nc99azuoFxo?si=p_iWkGqxCo5gEtdy

Here’s the original:

https://www.youtube.com/watch?v=iywaBOMvYLI

" I did feel the ‘composer’ using AI might have done a better job by making her increasingly complex prompt include timing information for hit points/mood changes"

I got that impression too, as if the AI was driving her and not her driving the AI.

[Martin Walker]

It's brilliant. Scary. I prefer the vocals on the soul version to the original.

Interesting!

I may be latching onto stuff I've read about how to distinguish AI music, but while the AI soul/jazz version sounded intriguing, I did notice that it had a soft/hazy/retro sound quality and lack of clarity that perhaps came from its training/stealing of lowres MP3 library content?

I've recorded the documentary for later viewing.

[BJG145]

…yes, it does sound lo-fi, which is OK in this context. If it’s got this good this fast, I expect it will learn to clean things up. It’s mind-boggling that it’s already so good. Wonder what it’ll be like in 20 years.

I think it was Folderol who previously suggested that AI output could be degraded by absorbing its own creations as training data. I have a hunch that it would be in the best interests of the AI creators as well as everyone else if they added a watermark which could be used to tell them apart...even if they kept it secret for now. Maybe they do; who knows. I doubt it though.

[Martin Walker]

…yes, it does sound lo-fi, which is OK in this context. If it’s got this good this fast, I expect it will learn to clean things up.

Possibly, but my point was that most music AIs seem to be trained on content pilfered from streaming libraries who rely on MP3 format - I wonder how likely it is that they will ever be able to improve the audio quality without access to hires content?

[BJG145]

...yep. I expect they will though. AIs don't just rearrange their training data; there are more dimensions to the processing than that. I don't claim to understand it, but I think it's pretty deep.

[Drew Stephenson]

Benn Jordan has been doing some work on this and has a tool that can identify AI generated tracks with a very high degree of accuracy because of the fact that they're trained on MP3s not WAVs. But I don't begin to understand the details.

[RichardT]

I assumed they were deliberately creating a retro feel with the AI. It sounded very much like a 70s recording, which I thought was appropriate.

The female vocal was more modern in style.

[BJG145]

...yes, it works in this context...but the question is whether they could have created a clean, modern version if they chose...and, I kind of doubt it, because I've noticed the lo-fi fuzziness that Martin points out in other AI tracks.

I think they'll figure it out though. For creators, things need to move in the direction of individual parts combined rather than a lump-mass-song. It'll be driven by economics, but I'm hoping there will be more control and separation...so the drums sound as if they might have come from a high-end sample library instead of a jumble of MP3s.

Essentially, I don't see why an AI couldn't learn the sound of clean jazz tracks by playing itself millions of them generated with the AI equivalent of a sample library, the same way they learned Go.

Who knows. If something's profitable enough, it's probably doable.

[James Perrett]

I did notice that it had a soft/hazy/retro sound quality and lack of clarity that perhaps came from its training/stealing of lowres MP3 library content?

Those aren't mp3 artefacts - it sounds more like worn vinyl played with a worn stylus. I wonder if the training prompt included the word vinyl?

[Martin Walker]

I did notice that it had a soft/hazy/retro sound quality and lack of clarity that perhaps came from its training/stealing of lowres MP3 library content?

Those aren't mp3 artefacts - it sounds more like worn vinyl played with a worn stylus. I wonder if the training prompt included the word vinyl?

Ah, could be - I did include the adjective 'retro' in my original post, and vinyl artefacts would be in line with that.

Which is actually more worrying, as I've noticed recent interest in slew limiting plugins that aim to add a certain 'analogue' quality to one's audio - if AI can already incorporate vinyl character into its sonic inventions then we're even more in jeopardy!

[ef37a]

I did notice that it had a soft/hazy/retro sound quality and lack of clarity that perhaps came from its training/stealing of lowres MP3 library content?

Those aren't mp3 artefacts - it sounds more like worn vinyl played with a worn stylus. I wonder if the training prompt included the word vinyl?

Could be! There are always "top" media people telling us that vinyl is the very best sound quality.

Dave.

[BJG145]

For creators, things need to move in the direction of individual parts combined rather than a lump-mass-song.

I just took another look at Suno. Things have moved on a bit since I last investigated this area, and it now offers a DAW-like interface.



Since Imgur went dark, I've been using Litterbox.

https://litterbox.catbox.moe/

You can host anything there. FWIW this is the track, generated from:

"Modern Blues, like Bonamassa, about an empty house"

https://litter.catbox.moe/bikskz40qoqwqg24.mp3

(I quite like the start, but it goes on a bit. It all looks controllable though; you can edit it in terms of length, sections, lyrics; upload your own audio etc. Looks like it can probably generate or extract MIDI as well.)

[Drew Stephenson]

Since Imgur went dark, I've been using Litterbox.

"This image failed to load."