Music AI: So?

 

I've been following the developments in music AI for a while. I'm still not impressed. The only AI music I have been impressed with to some degree is Emily Howell and Flow Machines, and they have plateaued. Generative music (SSEYO Koan) from the late 90s never went beyond the experimental. Brian Eno used it for a while for a systems approach but never caught on as a viable compositional methodology.

Facebook recently developed an AI that will take a small vocal sample or whistling and map it across a database of music samples. My first question was "why? or so?" (See the 5 Whys) Then I wondered how I could use it.

What a composer should want is a way to express a feeling or an idea. Even rock 'n' roll, which was a DIY "shortcut," still was an expression of a feeling or idea, not a science experiment. AI used in the arts should serve our humanistic desires, not smother them. I am a huge proponent of algorithms to facilitate problem-solving at the compositional level--not just in music, but in any activity that requires an examination of logic. Classical composers were extremely skilled at this, sometimes working by candlelight in extreme conditions for long hours. So it is regrettable to see AIs being built on samples of recordings of their works, so as to make other recordings more convenient and fecund--like cheap kitsch in place of fine craftsmanship. This is where the ethical boundaries arise: is it morally justified to use a classical composer's work, even if in the public domain, and especially if it is the pre-packaged sliced cold-cut version? 

Bot Bands

Databots is an AI metal "band." It is an interesting early crude prototype of AI music, that if allowed to follow the evolution of the talking head in TV, could perhaps have some relevance to musicians wanting to be on the vanguard.

Here are a few thoughts on their "Coditany of Timeness" album:

In "Memoryrian" (incidentally, all the titles are also generated by an algorithm if you haven't noticed), there are only two pitches that emerge from the noise, E and D and suggest E minor and D triads as a progression. Both humans and machines play by ear, but the human ear has eons of "machine learning", and understands why we have developed the concept and perception of "keys." A primary characteristic of metal is a blast of noise with manic drumming, and these are the first (and sometimes only) things the neural network recognizes. But how could it go deeper?

All pitched sound, regardless of what produces it, will want to resolve through the rules of equal temperament and into key centers. Listen to any one of the pieces by Databot, and it will create an earworm, and you could go to an instrument and pick out the pitches and find the key center. It's still the hard-wired "Newtonian" model of the musical universe; We are not easily soft-wired. "Software" has pervaded the last fifty years as a metaphor but it's still a hard problem. The software metaphor is driving AI music, then running on the old motherboards of human cognition.

There is an afterimage effect to noise: for example, if you play a piece of music, then follow it with white noise, a ghost of the music floats in and out of the noise, just as images did on tube TV sets. This is what the neural network "experiences." Artists have always noticed glitched imperfections in various media as a way to make art. But it is still esoteric, and if anything, would exist alongside other noise-based music.

Old M.O.'s

Recently, I stumbled on a documentary from the early 80s of Peter Gabriel when he first got into sampling, and it was interesting to observe the creative process in a time without PCs and the Internet. When you listen to all the rough takes by the musicians, compared to the released version, you are essentially hearing what AI is attempting to do. Even if machines provide some raw material, the creative decisions still have to be made by humans. (Can you imagine an AI that just uses all the bad takes?) What I'm interested in are ways of working, not necessarily having machines do it, or even assist with it. The big issue is efficiency: many people that have dishwashers seldom use them because they impose the idea of efficiency to the point it becomes inefficient. AI is that dishwasher in many ways.

Google Duplex, at least at the moment, is about useless automation and is in its very early talking head phase. Identifying what has been a tedious chore, typically leads to innovation. As far as I'm concerned, simple repetitive actions in an artistic context don't always have to be tedious. We marvel at things that are the culmination of thousands or millions of little actions, huge panels composed only of small elements like drinking straws or knotted gum wrappers, color-coded into representational mosaics, is not that same as the outputs of neural networks. It is evidence of how nature works until it evolves a heuristic. Google Duplex is a forced evolution and a solution in search of a problem.

The real problem for humans are the inefficiencies in multi-tasking. But like simple algorithms or macros, the results aren't always perfect, and still need sets of eyes.

Punished by Rewards

"Punished By Rewards" is a book released in 1993. The main premise is that excessive external rewards ultimately diminish motivation for improvement, and can sow cynicism and apathy. The efficiencies gained through technology have been rewarding, but we can also be punished or desensitized by those rewards. As in the example from 1982 in Peter Gabriel's studio, where we see him rummaging through a huge suitcase of cassettes with various recordings of the world music on them. There are no efficiencies desired because they weren't warranted. At the time, no one was complaining about how difficult or inefficient it was to go through a pile of cassettes, and in fact could sometimes be a pleasurable activity, such as we did making mixtapes for friends. Now we have the technology that replaced it, but we are punished by the rewards of convenience. Instead of 100 cassettes, we might have 2000 files, perhaps sequentially named: 0001, 0002, 0003. No wonder we don't care.

It's the efficiency paradox at play: in the pursuit of convenience and automation of everything, we throw the baby out with the bathwater. Hypergrowth and all its negative consequences are all born from the idea of efficiency. Planned obsolescence is based on the illusion of efficiency: the old one is always inefficient.

I am hopeful that AI can be integrated into creative workflows, but I'm not ready to say any art form is so inefficient that it should all be automated.  Making art manually is too much fun and spiritually rewarding.

Check out the excellent book re this: The Efficiency Paradox