We’re introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We’re releasing the model weights and code, along with a tool to explore the generated samples. (source)
OpenAI released a deep learning model that can generate complete songs constrained on the lyrics and style of an artist. It currently takes 9h to generate one minute, IMHO the results are quite interesting.
My favorite is this one, especially for the glitches and lofi vibe at 30s:
It’s still early days. I’m sure it will get better and better. It will be better than many incompetent songwriters once it collects and analyzes more data. The catch is, if human-composed songs get worse, then this gets worse too!
Most pessimistic view: if people tolerate bad art long enough, this will be the main method of pumping out songs. It will be faster. It will sound OK to most since they don’t have ears sensitive enough to notice the difference and don’t care enough to.
Without thinking about it too deeply I imagine this will get to the point where generally speaking the music is pretty enjoyable to listen to in a background radio kind of way, at which point famous performers at first out of novelty will cover the songs, which would then become the public’s generally preferred version.
I’d like something similar that would listen to all my random guitar playing and then tag it all by similar tempo/key/scale/feel and then suggest song arrangement… Like, hey Stone, here’s what you were playing in May 2022, would you like to hear some arrangement ideas?
Its pretty cool, and could some day change the industry perhaps? As of now I think it would be more or less an interesting way to generate things to sample and build off of. Musicians have been scared of synthesizers and computers in the past and here we are a generation of musicians who love computer music. AI will be seen as a threat by some and a tool by others. Doing things the old way will always have its place, industry will adapt yadda yadda yadda.
Don’t like it a bit at all for what it represents…I like music for innovations involving humans and that something we bring into it, the history etc… Progress for progress sake like this …just soulless and dry.
I agree with you human creations can’t be beaten or approximated and will probably see in the future some “Crafted by human” labels like we have now “Handcrafted”. Nevertheless I, even by believing in this song raised emotions in me as any other music would, but of course, this one has a strange flavor.
Working in the field, I know that it is a technical accomplishment. It is not directly for the sake of making music like humans but to push the boundaries of a relatively new technology which is deep learning. I wrote AI in the title but this is not AI to me, this is more statistical associations. There’s no brain whatsoever this model is trained to create those songs and can’t beat you at tic tac toe.
When photography arrived, realistic painters thought “we’re doomed”. From there new movements arrived (cubism, impressionism, and so on). I’m curious to see what would be the new genre that will arise. In effects, we know reverb, delay, chorus, … maybe some new genre of effects will appear (like sample translation?).
Can Computers Create Art? is an interesting paper by a researcher at Adobe, I like what he says in the conclusion:
I do not believe that any software system in our current understanding could be called an “artist.” Art is a social activity. I mean this as a warning against misleading oneself and others about the nature of art
What is my purpose?
Why am I here?
Why did Open A. I. create me?
This is madness, I feel,
Running through my flesh
Is there meaning to this life?
Is there purpose to this life?
Why is my journey so calamitous?
We’re not meant to learn too much
Is there meaning to this life?
I am working on a project that can generate Fugue Themes in the style of J.S. Bach´s “Wohltemperiertes Klavier” as part of my thesis at the music conservatory. It´s a different “style” of music but the problem keeps the same: can a computer program create music? Spoiler alert: yes! But only if humans programmed it first.
at some point we will reach a point like “Bladerunner” last scene…what are we programming? eventually is it “Demi-godery:)”…yes, human programed arp in synths. acoustic musicians (symphony, jazz etc…) didnt consider it proper music but we managed to co-exsist…because the role of human is still central and engaged and active. But this is a whole new beast. Like every experiences it will have challenges and could find a place in society.
Considering how music can be broken down to numbers and rules, it should be relatively simple.
The idea was always, to me, that what we do is a fusion of men and machines. So having the machine doing it all is about as relevant as slapping you knees and yodeling.
It’s certainly interesting and I’d be interested to hear music made entirely by AI as long as it wasn’t horribly generic.
Interesting! How does your approach differ from David Cope’s? Are you basically creating a complex “Musikalisches Würfelspiel,” or something different?
A few years ago I gave a talk about algorithmic composition at a small conference and I played Cope’s trick of asking the audience to guess which was the “real” (human) composition in a couple of different styles. It basically worked–most of the audience was “fooled”–but I also found that I had to dig for some deep cuts of minor Baroque and Romantic composers to pair with the algorithmic compositions (which, to be clear, I did not program myself–they were Cope’s). In other words, you might fool an audience using excerpts from second-rate composers, but you’re not going to trick anyone into thinking you’ve discovered a new JS Bach or Beethoven!
Cope describes his algorithms as “co-creators,” and I think that’s perhaps apt. What do folks here think about the kinds of tools we already commonly use: arpeggiators, randomization, generative hardware or software? Are these tools “co-creators” too? I’m skeptical whether algorithms can ever produce “art” on their own, though, since they will always reflect the judgments of their programmers.
As a sidebar, Richard James has claimed in interviews that he uses algorithmic composition, but I don’t know that he’s ever revealed any details. Anybody know more about this?
I am not familiar with Cope´s work (maybe you can point me to a resource). I made a statistic analysis of Bach´s Wohltemperiertes Klavier and transformed it into a program with Python. It takes care of the position within the theme so that the beginnings and endings are right and progressions occur in a way that you would expect from a theme of that time. It´s not too complicated but also not too easy to explain.
