The Quest for Musical Intelligence: (Part 1) The History
Before anything, take a listen to the music below:
https://www.youtube.com/watch?v=QEjdiE0AoCU
It’s pretty right? Music for a rainy day? A little repetitive, but the cascading piano figure adds melancholy to each slight variation, expressing a certain sadness that whomever wrote it must have intended to convey. Well, that “Whomever” is named Emily Howell, and is less a “who” than a “what.” That’s because Emily Howell can’t feel melancholy or anything as apparently irrelevant to music as that because she is a computer program. An Artificial Musical Intelligence.
Now, some people get scared about that. They might view art and music as an innately and uniquely “human” thing, which couldn’t possibly be replicated by zeros and ones. And I can relate certainly, after all, I was a composer myself before moving towards data science. But hearing the works created by these programs began to change my mind. Of course the process is not perfect yet.
And the programs, although growing and becoming more complex, are not perfect. That repetition from the Emily Howell piece is one of the hurdles of developing musical intelligence. Even if someone might be fooled at first, often listening to more of the repertoire betrays the author’s nature.
I intend to explore Musical Intelligence, from its histories to its processes, implementations and complications; I find this growing realm of study and creation truly fascinating, and I thought I might as well catalog that exploration for you to follow along.
Emily Howell was not the beginning, nor will she be the final creation in Musical intelligence. However, modern musical intelligence does start with Emily Howell’s progenitor. As Emily Howell births music, she was the brainchild of one composer turn computer scientist: David Cope.
David Cope and EMI:
Now one thing I was shocked to discover is just how old the field of Musical Intelligence is. It dates back to 1981, the same year as The Great Muppet Caper.
When I began my personal research into Musical Intelligence years ago, it was fittingly much in the same manner as David Cope did: composers block. David Cope was commissioned to write his first opera in 1981. A great honor, one that few composers of his age can lay claim to. However, any momentum Cope had was stopped in its tracks by a complete and total inability to write. Thus, he sought out alternate means of composing.
Living in the same area as IBM, Cope would begin to work with friends who worked there to construct a computer program that would be able to its own music based on a style.
But to do this, he questioned what musical style truly means. Cope once said in an interview, “I couldn’t figure out what my style was, so I suddenly started to say ‘well, I don’t know what my style is, so I should figure out what style is, period.’”
So his first program was more narrow in scope than a computer conjuring a masterpiece from nothing. Experiments in Musical Intelligence (or EMI (or Emmy)) was specifically designed to read in a set of a composer works, say Bach, or Brahms or Mahler. And, after learning that composers unique rules and idiosyncrasies, it would begin churning out music crafted with the tools of that style. (Mind you, this is back when punch cards were only mode of executing these programs, so churning might be a bit of an exaggeration in terms of speed.)
However, predictability became an issue. EMI would learn the rules and structure of a Bach chorale, but adhering to those rules so closely led to it sounding, well, robotic. To bypass this pitfall, David Cope created an unpredictability engine, that would insert calculated randomness, breaking the rules of whatever composer EMI was trying to emulate just enough to capture that chaotic “human” element.
Flow and adaptation, something many artists considered exclusive from machines had been recreated with zeros and ones. “[Cope found] multiple places where [Bach] broke his own rules, where he defied expectation of a particular progression. Cope developed “a little analytical engine” that could insert some randomness within the predictability” (Adams).
Below is an example of EMI’s take on Mahler’s works, compared to Mahler’s Symphony №1. Admittedly, when listening to both side by side, it’s pretty clear which one is computer generated, but the feat of getting this close is impressive enough.
https://www.youtube.com/watch?v=uVXYvhiG430
https://www.youtube.com/watch?v=_JXMFbGRyII&t=4m50s
And so, 7 years after being commissioned to write the opera, David Cope fed EMI his own works as if they were another Baroque composer, and once all of his own style had been metaphorically digested, EMI churned out an entire opera faster than a Jazz man could have improvised it. (Figuratively. We are talking about punch cards here, so it appears to have been tedious at every step).
To put this in a timeline, EMI was recreating Bach chorales before Caddyshack II came out.
The Next Stage: Emily Howell
We’ll never know if whoever commissioned the opera was happy about how long it took or the means by which it was delivered, but David Cope ended up expanding more on EMI. Making it more robust, and then eventually working on Emily Howell.
Emily Howell’s development began in 2003 (for time reference in terms of how old that is, that’s the same year as the Kim Possible Movie.)
And Emily’s process of creation was definitively divergent from her predecessor. Emily instead of learning the system and rules from pre-existing work, would develop her own style. The process involved human-computer interaction, with Cope teaching her mechanics and style.
“Every time you input a word, or a note, or a chord or anything you’d like to define as a unit, it is placed within a kind of virtual node. These nodes are placed in the order in which they’re received, and they’re connected to every other node with a series of arbitrary, randomized weights.” (Heater)
So what next?
From here, now that I’ve gone into the history see how it got to achievements like Emily Howell, I want to know what advancements they’ve made in the past 10 years, both through improved processes and implementations.
I intend to delve deeper into the techniques utilized, such as Markov chains as mentioned above and machine learning techniques, and hopefully under the hood mechanics that I don’t even understand yet.
In addition, the technology has become advanced enough to become marketable. There’s an entire industry built around AI services for creating music… IBM Watson Beat, Google Magenta’s NSynth Super, Jukedeck, Melodrive, Spotify’s Creator Technology Research Lab, and Amper Music.” (Deahl). I’m really looking forward to learning about each of these future programs that could yield the next breakthrough in music intelligence.
Maybe even get in to automated performance. I mean, just look at this little guy rocking out:
https://www.youtube.com/watch?time_continue=120&v=kG16-yIE8Sw
Conclusion:
It’s already come a long way, and musical intelligence has already begun wiggling its way into our culture. Perhaps machines have more of that “human” element we give them.
Resources:
Adams, Tim. “David Cope: ‘You Pushed the Button and out Came Hundreds and Thousands of Sonatas’.” The Guardian, Guardian News and Media, 10 July 2010, https://www.theguardian.com/technology/2010/jul/11/david-cope-computer-composer.
Heater, Brian. “Switched on Bach: David Cope’s Computer Compositions.” Engadget, 28 May 2013, www.engadget.com/2013/05/28/david-cope/.
Deahl, Dani. “How AI-Generated Music Is Changing the Way Hits Are Made.” The Verge, The Verge, 31 Aug. 2018, https://www.theverge.com/2018/8/31/17777008/artificial-intelligence-taryn-southern-amper-music.
