When it comes to music, composition is one problem that has always vexed computers. Sure, they can string a few notes or lyrics together, but creating motifs—repeated musical patterns that create form and meaning—is a real challenge. Now, after more than seven years of riffing on the problem, an Idaho State University Computer Scientist has come up with the solution.
“There’s this problem in the artificial intelligence field known as the long-range dependency problem,” explains Dr. Paul Bodily, Assistant Professor of Computer Science at ISU. “Basically, it’s really hard for AI to generate sequences of elements— like music or language—where what elements are chosen in one part of the sequence depend on what elements are chosen in a completely different part of the sequence. For example, musicians may want to have a chorus repeat in different places in a song or have lyrics at different positions rhyme. It is simple enough to copy and paste or separately generate interdependent sequences of notes. But typically, these approaches mean sacrificing the natural flow or coherency between the interdependent elements, and that’s where the challenge lies.”
. Pop* - pronounced ‘pop star’ - is an AI
system developed by Bodily
. It is designed to operate without needing a human and, most importantly, solves the long-range dependency problem.
“Pop* uses a special type of machine learning model that looks way ahead in the sequence at what will eventually be generated to adjust the probabilities of what it generates in the short run,” Bodily said. “In the end, the piece is coherent all the way through.”
Bodily used well-known songs like “Twinkle, Twinkle Little Star,” and “Somewhere Over the Rainbow,” as well as pop classics like “Hey Jude” by The Beatles and John Lennon’s “Imagine” to train Pop* on chords, pitches, rhythms, and lyrics. After this crash course training phase, the AI can now compose its own short pop songs. For lyrical inspiration, Pop* looks to Twitter for tweets on its favorite themes: being in love, feeling depressed, and new beginnings.
“One of the first compositions Pop* generated was this really somber, pensive song with the lyrics that went, ‘And I think I am just a lie,’ Bodily explained. “I was floored by how much it moved me. We humans all deal with feelings of inadequacy at times, but beyond that, here is an AI system that seems to be questioning its own validity and doing so in an artistic, creative way.”
“When I started doing research in AI, I was already heavily involved in a completely different research direction, and it was a big leap of faith to switch,” Bodily said. “This publication represents the reward for taking that leap seven years ago.”
In June, Pop* will be entering a battle of the bands of sorts, the AI Song Contest 2022
, and will be competing using one of its earliest compositions, “And I Think I.” Looking to the future, Bodily
aims to research the use of systems like Pop* for therapeutic purposes. He’ll be looking into how pairing people with an AI
that knows how to compose music
can help enhance their ability to be creative and process whatever challenges they may be going through.
“My vision is to integrate Pop* into a mobile app in which the user is prompted for thoughts or feelings, and then Pop* and the user work together to create a meaningful musical expression of or reaction to those ideas. The system could be designed to help the user share these creations or find other like-minded creators with whom they could connect. My hypothesis is that a tool like that could be quite powerful in helping people address various mental health challenges.”
For more information on the Idaho State University Department of Computer Science, visit isu.edu/cs