Musical Composition via Artificial Intelligence - A Status Report

I was just reading an article in The New York Times that was reporting researchers from the Massachusetts Institute of Technology, among other universities, have developed a new type of machine learning a.k.a. Artificial Intelligence (AI). The reason this is so newsworthy is that it is a vision program, meaning the information was fed into the computer program via a camera. Not only that, and here’s the kicker: It outperformed humans at correctly identifying handwritten characters based on a single example. That is incredible.

That got me to thinking: where are we with AI when it comes to musical composition?

Surely we must be further down the road, after all music is math and computers crush humans in that category. I have a sneaking suspicion I know what you are thinking already; music takes creativity and computers don’t have that yet! 

Well, let’s put it to the test.

The Turing Test

The Turing Test was named after Alan Turing. Considered by many to be the father of modern computing. He was also one of the first people to being seriously considering AI. His test for determining if a program was advanced enough to be considered intelligent was put forth in his paper “Computer Machinery and Intelligence”.

Without getting too far into what the test is, its strengths, its weaknesses, where and how it has implemented let me explain using music in a very basic example. All I would need to do is present a number of musical compositions to a group of people, presumably well-trained musicians, some of them written by AI and some by humans. The listeners would then decide which, AI or humans, had written each piece. If the pieces by AI were labeled as human by a rate above better than chance the programs that wrote them would be labeled as actual AI. Again, super simplified, but the concept is there.

Let’s try it out. I assume if you are here you consider yourself a musician of some sort, or at the very least an avid music listener. Here are two pieces. 

You have four options:

1. Both are Human
2. Both are AI
3. The top is AI and the bottom is Human
4. The bottom is Human and the top is AI

EXAMPLE 1:

EXAMPLE 2:

Answers are below in the grey box... no peaking now... ;-)

How did you do? Now, do you think today’s pop music is harder to write than classic orchestral arrangements? 

Experiments in Musical Intelligence 

The AI example is titled Zodiac and was algorithmically created back in 2012 by EMI or EMMY (Experiments in Musical Intelligence) which is a computer program created by David Cope. I highly suggest going to his site and reading about what went into EMI.

Cope has been gracious enough to put up 5,000 MIDI files of computer-created Bach-style chorales that you can go and download! 

Omax

Omax is another AI program from a few years ago that performs with us humans in a live situation. It actually learns real-time “typical” features of a musician and creates accompaniment on the fly! Yes, that is right! Your musical partner can be AI.  

But these are years old. 

Where are we today?

Jukedeck is the brainchild of researchers and developers from the University of Cambridge. It uses machine learning to create unique pieces of music based on two inputs and the time length desired. Seriously, you choose the two inputs and the AI makes a one-of-a-kind piece of music for you in less than a minute. 

Currently the options are *genre; styles:

• Folk; Uplifting, Melancholic 
• Rock; Uplifting, Dark, Angry
• Electronic; Uplifting, Corporate, Chilled, Aggressive
• Ambient; Sparse, Meditative, Sci-Fi

That might not seem like a lot, but the algorithm is able to make an endless catalog from just those inputs. Not only that, but the original pieces are created at jaw-dropping speed. A completely original piece clocking in at 2:30 is generated in less than a minute! Hopefully there are more options to come, but I am already blown away.

One of the great things about Jukedeck is that what is created is completely unique, royalty free and you get up to 5 downloads a month without paying anything!! There are other options for those that need more, but wow… I mean, just wow. 

This is the perfect option for budding video creators that can’t afford an original score! 

The Future

I think that the future is very bright. I don’t see any of this as a bad thing. 

Most main stream music is already comprised of the same chord structure—The I, V, vi, and IV chord progression. This was famously parodied by The Axis of Awesome in their song “4 Chords”.  In the music video on YouTube the band pays 47 chart topping pop hits that all use the I, V, vi, IV progression. Their live shows have included up to 27 additional songs as well.

Not only that but the two most popular scales (keys) used in a database that contains the analysis results of 13,000 pop songs are C/Amin (26%) and G/Emin (12%) for a total of 38%! This is most likely due to playability. C/Amin contain no sharps and G/Emin only contain 1. 

I would seem these statistics are revealing something more robotic than what is being done with AI. Not only that but AI will treat more complicated scales and progressions just the same as the easy ones, because math is math. 

I am looking forward to hearing the future! AI will help us all push the boundaries of our own music and music in general. So to borrow from Dylan Thomas... Don’t go QUIETLY into that good night, and do not fright—AI is here to help!

Learn music theory in The Ask.Academy here.