AI Generated Music
-
@Arantor said in I, ChatGPT:
Someone on Twitter, I forget who, nor do I care, was very excited about this.
-
@sockpuppet7 I remember a Google Doodle that did AI generation of music "in the style of Bach" based on a bit of melody you could supply. It was entertaining... but not very good.
Would it be better now? Well, maybe. The core issue is that there really isn't all that much music by any one composer, so learning a style is difficult with current LLM methods.
-
@dkf said in AI Generated Music:
@sockpuppet7 I remember a Google Doodle that did AI generation of music "in the style of Bach" based on a bit of melody you could supply. It was entertaining... but not very good.
Would it be better now? Well, maybe. The core issue is that there really isn't all that much music by any one composer, so learning a style is difficult with current LLM methods.
This one I found hard to distinguish from music created by humans
-
For some extra irony.
-
This one is about Orko the Great
-
Some of these are very cromulent.
-
I really liked this one, although it seems to consist solely of an outro.
-
@Zecc reminds me of GlaDOS with the intonations being just slightly stilted and not quite to the meter or rhythm properly.
-
-
@dkf said in AI Generated Music:
The core issue is that there really isn't all that much music by any one composer, so learning a style is difficult with current LLM methods.
That depends on which composer. Some composers wrote hundreds of pieces; Telemann wrote over a thousand. That's small compared to the countless texts available to train an LLM, but as large or larger than the number of texts by any specific author. If an LLM can learn the style of a particular author from the relatively small number of works by that specific author, I don't see why similar methods couldn't be used to learn musical styles.
A bigger problem might be similarity of styles. Certain stylistic traits tend to be characteristic of an era and common to all composers of that period. Music of Bach sounds, at least superficially, quite similar to Handel, Vivaldi, Telemann, Buxtehude, and other contemporaries. There are recognizable differences, but they're subtle compared to the similarities. Likewise Mozart, Haydn, and their contemporaries.
I don't know squat about how LLMs work, but it doesn't seem to me that the corpus of musical training material is fundamentally different from the corpus of literary training material. There is a vast amount of both available, with rather small amounts from any specific individual.
-
@Zecc said in AI Generated Music:
Some of these are very cromulent.
That sounds better than anything Lil' Shit on Yo Face feat. the Subways has released the last few years.
Maybe music artists actually do have something to fear.
-
@topspin said in AI Generated Music:
@Zecc said in AI Generated Music:
Some of these are very cromulent.
That sounds better than anything Lil' Shit on Yo Face feat. the Subways has released the last few years.
Maybe music artists actually do have something to fear.Subways had one good song and coasted on it for far too long. They arenât half bad live but I wouldnât fork out to see them at current prices.
On-topic: Iâll say what I said before. 99% of people arenât interested in new or quirky things. They find a thing they like and stick to it. LLM derivatives will give that to people in spades and will probably eat everything from the lower end to middle-upper end of any market theyâre deployed in. Prepare for garbage.
-
@HardwareGeek said in AI Generated Music:
Some composers wrote hundreds of pieces; Telemann wrote over a thousand. That's small compared to the countless texts available to train an LLM, but as large or larger than the number of texts by any specific author. If an LLM can learn the style of a particular author from the relatively small number of works by that specific author, I don't see why similar methods couldn't be used to learn musical styles.
I believe the list of Bach's works goes to over 800. The problem is that current standard AI learning methods need many orders of magnitude more examples than that to converge.
This is a huge problem in many fields, and one of the key things holding back AI development; a great many of the other problems stem from it. For example, the problems with slurping in data from people who didn't explicitly agree to that use all come from the need to Feed The Maw, as do all the issues to do regression towards the mediocre (via the use of simpler AIs to generate training datasets) and the looming elephant in the room of energy consumption.
The right thing would be to throw significant effort at developing a better learning algorithm (there's very good reason to believe they exist; there are some academic demonstrators) to unlock the potential for far more interesting applications where the training data is sparser. That is not the Silicon Valley Way however, and they're doing their level best to suck the oxygen away from the alternatives.
-
I swear yesterday Suno's Explore page was just listing some trending songs people created; which is from where I got the ones I post above.
But today it just gives me a bunch of clickable random genre mashups and when I pick one it plays music from that genre.
I'm currently using it as a background noise generator in a genre that I find works for me (fwiw, named "african folk math rock", though it doesn't sound very math-rocky to me) and I'm finding it to be very effective for the most part.
-
@dkf said in AI Generated Music:
I believe the list of Bach's works goes to over 800. The problem is that current standard AI learning methods need many orders of magnitude more examples than that to converge.
Worse, the recent reports show that for linear increases in quality, logarithmically more data is needed.
-
It change the genre I was listening to on its own.
I wasn't expecting French Reggaeton.
-
A recipe for spaghetti to music
edit: figured out how to link the song itself
-
@Arantor said in AI Generated Music:
Worse, the recent reports show that for linear increases in quality,
logarithmicallyexponentially more data is needed.FTF what the reports actually said. A logarithmic increase in data for a linear increase in quality would be superb.
-
Won't let you make sexually explicit songs.
-
@loopback0 was expecting 8 Mile. Disappoint.
-
The true cost of rockâs multi-million-pound hologram gamble
Abbaâs phenomenally successful virtual show has created a revolution in concert-going â but some fear for the worst
Somewhat related. I can actually see more of these turning up. Once they're cheap and mobile, we're going to be infested with them and we might see the death of the coverband.
If you go to a Kiss concert, it is families, everyone is there â all generations. It is very important to have a [broad] fanbase
I don't think you've been to many Kiss concerts. They're mostly populated by aging rockers with sharp elbows.
Although the likes of Swift and BeyoncĂŠ are breaking records with their supersized arena tours, tickets for smaller gigs are going unsold and 125 grassroots music venues shut last year, according to the Music Venue Trust.
That's because the smaller venues are charging almost as much as the stadium gigs now. Are you going to risk ÂŁ60 a go on an up and coming band or risk ÂŁ100 on a stadium gig with a band you love? Both can be equally shit but at least with the band you love, you can sing along.
*edit you know what's funny about Kiss though. They've been taking potshots at other acts for miming and that bump note substitution for a couple decades now. This would be hilarious if they were first.
-
@DogsB said in AI Generated Music:
Are you going to risk ÂŁ60 a go on an up and coming band
Might have to eat crow on that. Just got a Kris Barras ticket for ÂŁ31.
