music and artificial intelligence

Hi everyone .
There was a story on the news recently that I think is known to most of the world's population.
The last Beatles song that has never been published before has found new life thanks to artificial intelligence. They managed to reconstruct a ruined song using other Beatles songs.
I wonder if it's possible to do the same thing with old, poorly made recordings. I'm not referring to the Beatles but to any other musical group that wasn't lucky enough to record their music with decent equipment.
 
Member
Joined 2015
Paid Member
There are already papers on Arxiv about generative music restoration utilizing diffusion models trained on a curated set of music tracks, see as example 2403.18636. It is likely that commercial applications will follow as soon as acceptable quality will be reached.
 
Hi everyone .
There was a story on the news recently that I think is known to most of the world's population.
The last Beatles song that has never been published before has found new life thanks to artificial intelligence. They managed to reconstruct a ruined song using other Beatles songs.
I wonder if it's possible to do the same thing with old, poorly made recordings. I'm not referring to the Beatles but to any other musical group that wasn't lucky enough to record their music with decent equipment.
According to Paul and Ringo, the song was not 'constructed' using AI: AI was used to separate John's vocal from his demo so that it was usable. George had done a basic rhythm track at the time, so that was used. Paul and Ringo then finished the track, with (I think) some background vocals from an old Beatles song.

Apart from the use of AI to do that, the song was made using the same techniques used for the past 60 years or so: not all of the musicians are in a studio at the same time; tracks or bars are cut and pasted from various sources and a finished track is often, or most of the time, made from multiple sources.

There were of course many Beatles songs where only two or three members - or even just one - were on the recording, or a basic track was overdubbed later.

If it could be done, and if the Estates agreed, I'd love to hear what could be done with the 1962 Star Club tapes: poor sound but great energy, bum notes and atmosphere.

Geoff
 
re: use of STEM generators -can you think of simple workarounds to get better result with the "BASS" track? - - most STEM category work but "bass" is typically decimated with no transient info - overtones are gone (Rip-X Spleeter) - leaving a muddy - "phasey" sounding characterless mess.
 
Further to post #3, prior to the official release of "Now and Then", some people seem to have gained access to the demo, added instrumentation and uploaded onto YouTube. This was a few of years ago before AI software had been fully developed.

Being selfish about it, I'd love to hear what AI could do do some of my favourite bootlegs and old recordings such as Charlie Christian, as long as the charm of the original was maintained and nothing new was added.

For example, Hendrix' marvellous 'Live at the LA Forum 1970' show, which only exists as a low fi mono audience recording, could provide an interesting basis for such an experiment: all the instruments except the bass drums (Mitch used two) are at least audible. Voice and guitar are OK considering the origin of the recording. It would never be hi fi of course.

Geoff