Real Musicians Evaluate Music Made by Artificial Intelligence

This Irish folk algorithm is a surprisingly good original composer.
June 18, 2017, 1:00pm
Image: Shutterstock

Artificial intelligence has already proven its worth at performing routine tasks. But how good can it be at artistic creativity? A group of researchers at Kingston University and Queen Mary University of London are exploring the concept with AI that composes music.

folk-rnn, the machine learning algorithm developed by the researchers, was trained with thousands of samples to produce new Irish folkloric tunes. According to Dr. Oded Ben-Tal, lead researcher and senior lecturer in music technology at Kingston University, the results are impressive and can mark a new chapter in music, in which humans and machines collaborate in creating art.

There are practical reasons to start with Irish music, Ben-Tal said, including the availability of data. "In order to use machine learning methods to create useful models, we need lots of data," he said in an email to Motherboard, which he cowrote with his collaborator, Dr. Bob Sturm, digital media lecturer at Queen Mary University.

For their research, the team used a large online crowd-sourced repertoire of 23,000 folk music transcripts, recorded in ABC notation, a shorthand form of representing music by using the letters A-G and symbol.


From a technical perspective, Celtic music is much easier to model with computers, Sturm explained. "The collection of tunes we are using share a consistency of structures at many different time scales, which really motivates the application of statistical modeling methods," he said.

The initial version of folk-rnn, released in 2015, functioned like other natural language processing models, parsing ABC transcriptions like natural text and finding patterns, which it then uses to create original melodies. Since then, the algorithm has undergone two more iterations to become more efficient.

"I was surprised to see that there were quite a few tunes which had an Irish feel to them."

Ben-Tal and Sturm have published 3,000 transcriptions generated by their second model under the title "The folk-rnn Session Book Volume 1 of 10." Synthesized versions of tunes created by different iterations of the algorithm can be accessed at "The Endless folk-rnn Traditional Music Session."

The quality of the AI-composed music can only be appraised by human musicians—at least for the time being. "In the UK, we are surrounded by many practitioners of this kind of music," Ben-Tal and Sturm wrote in their letter. "We can thus involve these experts in our research to more meaningfully evaluate the behaviour of our trained systems, and their use in music practice."

Motherboard reached out to Irish musician and composer Daren Banarsë, who had the chance to examine the first volume of folk-rnn's compositions. "I was surprised to see that there were quite a few tunes which had an Irish feel to them," Banarsë said in an email to Motherboard. "Most of them would need some human input, some reworking, to sound convincing. But there were a few that were performance ready, or just needed some rephrasing, or a couple of notes changed."

He did point out however, that performance-ready doesn't mean that a computer synthesizer would be able to play the tune convincingly. "It's worth noting that with Irish music, the score is taken as a guide, a skeleton which is fleshed out with the musicians' ornamentation and personal style," Banarsë said. "We add the style, the swing, the energy and ornamentation. If you were to hear the computer playing the tune without any human intervention, it would sound crude and lifeless, and probably not recognisable as Irish."

One of the interesting aspects of computer-generated music, Banarsë said, was the "mistakes" that the algorithm had made, which could bring up new ideas. "There was one reel which intrigued me. The melody kept oscillating between major and minor, in a somewhat random fashion. Stylistically, it was incorrect, but it was quirky, something I wouldn't have thought of myself," he said.


In the past years, a handful of companies and research labs have tried to venture in the domain of composing and playing music, including Google's Magenta Project, which aims to use machines to "create compelling art and music." And as with everything related to AI, there's fear of algorithms driving humans into unemployment.

So will Ben-Tal and Sturm's work put Irish musicians out of work? The researchers dismissed such a notion.

"Our work does not seek to replace or best humans," they wrote. "We are interested in making artificial systems that can augment human music creativity in helpful ways. We aim to produce useful artificial collaborators that make intelligent suggestions, and open up new avenues for creating music."

"As a composer, I have no worries about my job being replaced in the near future," Banarsë said. "But as the technology progresses, who knows? I hope there'll be a stage before that, where the computer can assist me in my job. I always find it daunting when I have to start a large scale composition. Maybe I could give the computer a few parameters: the number of players, the mood, even the names of some of my favourite composers, and it could generate a basic structure for me. I wouldn't expect it to work out of the box, but it would be a starting point."

Get six of our favorite Motherboard stories every day by signing up for our newsletter .