Originally Posted by etcjoe
Revisiting this topic. After doing a load of research and speaking with our AI guys at work I have a better understanding. Suno and others like it are not regurgitating others music, or sampling it or copying it or any of that.
In some cases, it is "regurgitating other's music" - or at least portions of other people's copyright material.

For example, it's been shown to generate producer's tags - basically audio signatures placed at the start of tracks.

Quote
It was trained, supposedly on copyrighted music and based on all that training when given a prompt it actually creates a unique output that meets what it knows about that genre and any other qualifiers you may have used.
Those "qualifiers" are how it categorizes elements of the songs that it's learned. Using the right qualifiers, it's possible to make Suno create output that is quite similar to copyright materials.

Given that Suno is able to generate copyright-infringing output, it's clear that no "supposedly" is needed. Sunu is being trained on copyrighted material for which they have no license to use.

AI companies attempt to stop this by preventing users from using certain descriptors, but with some clever prompt engineering people have been able to get around that restriction.

Quote
So in summary Suno and others are not just combining pieces of other music and calling it new.
Yes and no.

It's not "combining" in the sense that it takes the first few seconds from one song, and then a couple more seconds from another song.

But it is combining in the sense that it's learning songs, and then creating output by morphing elements of those learned songs.

It's certainly more sophisticated that cutting and pasting between songs. But without having learned those songs - and stored them in its neural network - it would be unable to create new songs.

Quote
It generates from scratch a song, using programming of course, that mimics what it has learned that song should sound like, what instruments it should have, how that should sound, and what they should play given a chord structure it created as well.
I'm not sure what "using programming" means, but it doesn't really generate a song "from scratch". It generates a song by morphing together learned elements of songs from the training set. It just happens to have a huge amount of songs to draw from, so it's often difficult to see where it's come from.

But choose enough descriptors that match the training data, and it'll spit out elements of those songs, showing that it's just copying after all.

-- David Cuny
My virtual singer development blog

Vocal control, you say. Never heard of it. Is that some kind of ProTools thing?