Automatic music : is it possible ?

[Functional]

The problem with "automatic music" is that as much as you can devise an algorithm capable of making it, it's a whole other story whenever that algorithm is capable of telling a bigger story that hasn't been told yet. It is bound to either stay safely within a single expression that it keeps reiterating or it ends up with bunch of different expressions that are loosely connected.

I don't know how to explain it exactly in terms of music, but one explanation could be used through videogames. It's anything but unfeasible to make an algorithm that creates 2D puzzle games with some narrative. But good luck trying to get it to stitch together mechanics and storytelling so that they would compliment each other and be interesting.

This is especially true if you're going to use probabilities to do things like keychanges. The difference between going a halfstep or a fourthstep up is so drastic that you would never want to have a situation where probability distribution solves which one of the two it would be

[Tj Shredder]

It can be done without neural networks. Been there, done that. Well, not with music, but the concept could work.

At work we had fake test data of personal records with random lettered names. Garbage, unpronouncable.

So from the real data I counted how many surnames start with A, B, C etc. How many times is an A followed by an A, B or C etc?
Do that for all combinations of two letters, what is the chance its followed by whatever third.
This way you build up random names on probability, which turn out to be easy to pronounce.

Same thing can be done with musical intervals to build random melodies that are not conplete junk. Usind just a matrix, no network.

Yes, it is far from cognitive...

Markov chains! If you do that with Bach, you get something which sounds like Bach, but is dead boring. If you train a neural net you can get much more interesting results, but Bach already made more works than a normal human can hear during her life. There is the DeepBach project generating +5000 chorales in the style of Bach...

I do see a use case where a performer still would do the crucial decisions while performing. AI generated music has no meaning in itself, it needs human expression to add that. It could be just an interpretation of a computer generated score as minimum...

Jean Michel Jarres iOS app ION seems to use this technology:
http://www.flow-machines.com
I find it quite impressive...

[JackyZhylinska]

AUTO-ADMIN: Non-MP3, WAV, OGG, SoundCloud, YouTube, Vimeo, Twitter and Facebook links in this post have been protected automatically. Once the member reaches 5 posts the links will function as normal.

Sure! There are plenty apps which will help help you creating a unique piece based on algorithms. From my experience I would recommend this chord progression generator https://chordchord.com/ (https://chordchord.com/). I used it to write a couple of songs for my band, loved the result. My piano part became more interesting and emotional. So I think it's quite handy and cool to spice things up in your music piece with help of the programs.

[jancivil]

spamtastic

[cthonophonic]

Indeed, when so much marketing copy is written by people who struggle to sound convincingly human, I don't have a lot of faith that an algorithm is going to produce any music worth listening to.

Spammers these days. No commitment to the craft...

[vurt]

Indeed, when so much marketing copy is written by people who struggle to sound convincingly human, I don't have a lot of faith that an algorithm is going to produce any music worth listening to.

Spammers these days. No commitment to the craft...

maybe the spam is auto generated, and is therefore an example of the algorithms in a different medium, therefore showing its usefulness in many situations.

[deastman]

I think machine learning is more likely to yield convincing results than utilizing exclusively procedural methods. Of course, the output is entirely dependent on the training data set, which means that it can only parrot back what humans have already created. Then again, all of western music is based on arbitrary human constructs. (Not entirely sure what my point is here! Lol )

I find generative and procedural tools to be at their most compelling when they augment and collaborate with an artist. Art-directed proceduralism can save the artist from focusing on the smallest, most tedious, highest frequency details, while allowing them to execute their unique creative vision in a broader sense. Even better is when the results are in a form which allow the artist to edit and revise (e.g. midi data). Conversely, the less influence an artist has on the creative output, the less ownership they can claim. In a fully automatic scenario, the real art is not the output, but rather the code itself, and the developer is the true artist.

As for the examples posted so far, I find them intriguing and challenging. At times, they do sound like “real” music, but then the choices made are often ones I find unpleasant and what I would categorize as “bad” music created by someone with poor taste and a fundamental lack of compositional training.

[jancivil]

the OP has some pretty interesting material up at the youtube, I'm sure I don't follow his thinking from that post tho

I agree with all of that. I'm not positively disposed to 'exclusively procedural', the idea is tiresome. I wouldn't set a high bar for it, let's just say.
I like setting things up to a certain extent to give me unpredictable, and see what I can learn from something, as in relationships in the abstract

[deastman]

I’ve long been a fan of generative compositional techniques, to the extent that they provide something for me to respond to and to play against. Eventually I realized that I’m just mildly antisocial and don’t have much opportunity to play with other people. These tools offer me a collaborator of sorts.

[Timfonie]

Yes, automatic music can be done in truely every possible way. Just don't estimate what will ultimately be possible by judging only the current status and output of AI generated music.

The easier aspect for AI to master are things like: melody, rhythm, harmony and variable amounts of variation

A little harder but still attainable by AI these days are different styles and current trends, lyrics and overall structure.

The hardest part is creating something which fits current reality. It's far from easy for AI to come up with music including lyrics which somehow reflects on the current moods of society (and subsocieties).

I've spoken to experts on AI and Twitterbots. They expect AI to be able to write complete novels in about 9 years from now.

Ultimately AI will be used to create an abundance of music out of which only the very best are selected. Those may become popular. Which ones? That may still be a matter of taste and thus a matter of selection by humans. The point is: will this selection be done by by musicians, companies and DJ's, or will they be selected by the general public, by the listeners?

[pekbro]

I thought this was about something like automatic writing, only you would channel Beethoven's ghost or something to compose your music for you. *I am somewhat disappointed actually.

[Timfonie]

Yes, the OP topic is on algorithmic or mathematical generation of harmonies and melodies. I consider the examples given on Youtube convincing, in that they are more interesting and complex than most of us can compose without any aids.

I just took the freedom to extrapolate on the algorithmic generation of music in other aspects of music as well.

Why and in what way are you disappointed?

[pekbro]

I'm just more up for a psychic music discussion, than an artificial intelligence discussion.
Anyway, AI isn't really a big deal, they just need to come up with data storage facilities that
rival our own. Then a capable AI will likely come into existence on it's own eventually, despite
our interference.

[InLight-Tone]

It's a waste of time. Trust me been there. Compositional assistance sure, but that could be an Arpeggiator and other REALTIME processes. Spend your time learning composition. Mapping fractals to tones and other assorted methods yield random results with little patterning and cohesion. Humans relate to ORGANIZATION, NOT RANDOM!

"The likelihood of new material agreeing in style with the style of already composed music is also negligible." Cope The Algorithmic Composer

[BertKoor]

Markov chains!

Oh bummer, too late for a patent

Nice to know there's a name for the methodology. Next time I tell that story about fake surnames I can just say "so we used Markov chains"