Stats and Beethoven...

[PierreV]

Funny coincidence with the statistical thread, Google Now told me to look at this...
I can see advanced generative models showing up in the next few years... (as they did with text, some of them not fully explained given their potential for abuse)
I wonder how I would react to new pieces by a fully believable Beethoven.

https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0217242

 

[PaulD]

Hi Pierre, "fully believable Beethoven" won't happen. However, "able to fool the general public or reporters Beethoven" has already been done. I did my PhD in Algorithmic Composition. Even David Cope who has made a career out of promoting these sorts of ideas (see Experiments in Musical Intelligence) has not addressed what is called the Theory Problem, viz:
Composition is more than an application of music theory, as a creative process it breaks rules and makes rules at will to suit its needs. Computers cannot do that yet.

Music theory is created by analysing a bunch of compositions and finding common rules, ignoring the anomolies, thus it is a reductive process. Composition is an expansive process, and it a computer applies music theory rules to create music, then the result is average, because of the averaging of the rule set. The way to overcome this is to use controlled randomness, a la Xenakis or Koenig, to create compelling and sophisticated compositions.

The first book on algorithmic composition (Experimental Music by Hiller and Isaacson) made this observation in 1959. Modern attempts at using artificial intelligence have so far not advanced the art beyond controlled randomness. Some of the best examples of "style replication" come from using Markov chains for conditional generation, the way this works is to take a composition by, say, Beethoven and use random jumps into it such that (this is a second-order Markov chan example), I already have the notes "C" and "F", then I jump randomly into the Beethoven piece, and then step through it until I find a "C" followed by an "F", and I take the next note as the note I'm looking for in my piece. This will create a fairly convincing replication of Beethoven's style.

Google has been playing with its AI engine in this regard, but so far the results that I've heard are no better than Markov chains, although the way of getting to that result is fundamentally different.

I hope this helps!

 

[Xulonn]

Wow! Fascinating - and basically understandable by laypersons. ASR continues to amaze me with its collection of intelligent, educated participants who walk the line between science and the arts -the subjective and the objective - and strive to comprehend the relationships between them.

 

[Daverz]

 

[PierreV]

Google has been playing with its AI engine in this regard, but so far the results that I've heard are no better than Markov chains, although the way of getting to that result is fundamentally different.
I hope this helps!

Yes, thanks. I've heard pseudo Mozart based on Markov chains and yes, it was indeed fun but not convinving. I guess the GPT-2 results for text aren't going to be replicated soon for music, possibly because the corpus of a specific composer is a too small dataset compared to the somewhat formulaic human writing one, but I have been so impressed by the Go research that I am, I guess, wrongly impressed/pessimistic.

 

[PaulD]

I would not say you were "wrongly impressed/pessimistic", unless you are a professional 'classical' musician. Cope's examples of style replication are very impressive for many people and it takes trained listeners to hear the anomalies. It's a bit like many things, I'm sure I'd be fooled by things in your profession that you would see as glaringly obvious. I'm happy to point out texts, papers and resources if you are interested in going further.

 

[PaulD]

Wow! Fascinating - and basically understandable by laypersons. ASR continues to amaze me with its collection of intelligent, educated participants who walk the line between science and the arts -the subjective and the objective - and strive to comprehend the relationships between them.

Many thanks Xulonn, I'm glad it was understandable! I have appreciated your posts in the TotalFiasco DAC thread

 

[PaulD]

One other thing I would point out is that writing "good" music (I'd say "compelling") algorithmically is at least as difficult at just sitting down and banging it out at a piano and writing down the things you like. It is not just like "pressing a button", but rather it is having an over-arching idea, and then developing a system that takes the idea(s) and implements this in a way to output musical data. It often involves at least as much imagination and skill as any other type of composition. I'd also say that style replication is the least interesting area of algorithmic composition, but obviously fascinating from a stats viewpoint a la the article in the first post. Sorry, I obviously could go on about this ad nauseam... I'll shut up now

 

[SIY]

I'm happy to point out texts, papers and resources if you are interested in going further.

Yes! Are there any good introductory texts?

 

[PaulD]

There's the usual sort of things at Wikipedia here https://en.wikipedia.org/wiki/Algorithmic_composition

I very much like the introduction by Michael Edwards here here (PDF). or https://cacm.acm.org/magazines/2011/7/109891-algorithmic-composition/fulltext.

Martin Supper has a very good article here http://muse.jhu.edu/article/7797

This may be interesting and useful http://www.algorithmiccomposer.com

Full disclosure, I wrote a fairly critical review of Nierhause's book on Algorithmic Composition here. I think it really should have been titled Computational Musicology, or Computational Style Replication...

Hope that helps!

 

[pierre]

You can experiment with Magenta too. If you know some python and tensorflow it’s easy to create your own. It is fully open source.

 

[PaulD]

There are a LOT of open source tools for this sort of thing, from PD (Pure Data) to SuperCollider, CSound, and PatchWork, plus others, to free tools like AC Toolbox and others and paid things such as MaxMSP, Audulus and so on. Ableton is not so useful as it imposes assumptions of form and style on the process... Many of these things use MIDI so are limited to more standard ideas of tuning, but others are more flexible in that regard.

 

[PaulD]

Google created a 'doodle' for Bach's birthday here https://www.google.com/doodles/celebrating-johann-sebastian-bach. I wrote a review of it for a journal, but I found the engine behind it a little less of a toy and it's something you can play with in a browser here https://coconet.glitch.me - it's worth having a play with for a bit of fun.

To give you my take on it, the last two paragraphs of the review are:
I have worked with many algorithmic methods and I cannot say that the Google Doodle was noticeably better than any other. Certainly Markov models have been used very successfully for style replication before, by Clarence Barlow to cite just one example. Also, I did not find the harmonization from the Google engine to be particularly Bach-esque enough to be convincing. The added parts often showed poor voice leading and would not be easy to sing, like too much emphasis was placed on the vertical arrangement of the notes. I also thought the harmony was somewhat more simplistic than Bach would use. I did notice that it gave a more Bach-like result if the provided melody was more like Bach, rather than some random scattering of notes. I did put in a couple of bars of a Bach chorale and the results from the Doodle were quite different than the original, on each successive harmonization. Given how well-studied Bach’s music is, and how it seems more based on clear order and structures and rules than most music, I am surprised how poorly the AI engine does in recreating Bach’s style, and perhaps this just proves the gulf between human thought and creativity and our current AI models of it.

After playing with the Bach Doodle, I do not think it is sophisticated or extensive enough to use any other way than for play, as the limitations on pitch, length, rhythmic units and so on are too severe. However, as Jean Piaget said, “play is the work of children,” and perhaps compositional or analytical “cognitive development” can come from playing with simple generators such as Coucou.