It's All About The Algorithm, Baby!

by Joe Lavin

Exactly how predictable are we when it comes to music? Could a simple computer program predict the songs we’ll love? I prefer to think not, but, regardless of whether or not we like it, artificial intelligence is increasingly being used to predict our taste in music.

One Spanish company, Polyphonic HMI (www.polyphonichmi.com), has created Hit Song Science, an application that can take any song and break it down to its mathematical core. If you’re like me, you probably didn’t even know that pop music has a mathematical core, but it does, and apparently this can be used to determine a song’s “hitability.”

This software will compare the song to a database that contains the “top-30” hit songs of the past five years in order to search for mathematical similarities. The algorithm then assigns each song a score between one and 10. Any song rated more than seven is likely to become a hit. Fewer than four, and the musicians might want to keep their day jobs.

On the surface, the idea seems ludicrous, as if Deep Blue, the chess-playing computer, had some younger, underachieving brother who was really into the music scene. But many major record labels are already using this software. So far, the company claims to have predicted several hits, including those of Grammy award winner Norah Jones. At the time, her music sounded like nothing that was being played on the radio, and yet her songs still rated highly. Her album, Come Away With Me, has since sold more than 6 million copies, much to the surprise of many who doubted that a jazz singer could attract such mass appeal.

Polyphonic HMI has even released a scaled-down version of the program at www.hitsongscience.com, where aspiring songwriters can have MP3s of their songs analyzed for $49.99 per song. These reports are not as thorough as the ones music labels receive, but they do give a new band the chance to see how viable its dreams really are.

The technology was first developed when researchers at Group AIA, the parent company of Polyphonic HMI, decided to analyze the entire music universe. The company took the approximately 3.5 million songs that have been released commercially since the late 1950s and ran them through a computer.

“When we plotted all the songs on a map, it looked like the Milky Way system,” said Tracie Reed, vice president of North American operations for Polyphonic HMI. Later, the company stripped out just the songs that had been top-30 hits in the past five years. “We realized that our beautiful Milky Way no longer looked like the Milky Way, but like a series of constellations,” she said.

Within those constellations, the researchers noticed distinct patterns for hit music and found that a popular song could be grouped within one of 18 clusters. Songs within a cluster are similar mathematically, though they won’t necessarily sound similar. In fact, it could theoretically be possible for Enya and Metallica to be in the same cluster, though they obviously sound nothing alike. However, even if you do like wildly divergent sounds, there’s a good chance that your favorite songs will belong to the same cluster.

Along with discovering new hits, the company can also use its techniques to help consumers find new music to enjoy. Researchers are now developing a recommendation engine that will consult the company’s map of 3.5 million songs.

Currently, most recommendation programs are based on the purchases of other people. If you buy a Radiohead album at an online retailer, and another person who bought a Radiohead album also buys Cher’s Greatest Hits, then the site will probably recommend that you too buy Cher. Polyphonic HMI’s application instead will look at songs that are mathematically similar to Radiohead and recommend those. “It’s the purest recommendation application ever because it’s based on what you like,” Reed said. She is hoping that it will soon be available to consumers through music retailers.

Of course, all this math is a far cry from the glamour of rock ’n’ roll, and there are obvious complaints. Music on the radio already sounds the same. Won’t this just encourage more cookie-cutter pop songs? Actually, the company believes this software could have the opposite effect.

“This application is giving people the confidence to take more risks,” Reed explained. Far from making music more homogeneous, the company believes it could convince labels to go out on a limb more. If a bluegrass song that a music executive likes gets a score of more than seven, then that executive might be more willing to take a chance on it.

Still, it’s worrisome to think that computers are being used to determine our tastes. To be fair, the company doesn’t expect the software to replace music executives, though in some cases that might not be a bad idea. And it has no illusions that its system is perfect. After all, there is no way for the program to measure the lyrics of a song or the image of its performers. And it has no hormones, either. If Christina Aguilera is wearing practically nothing in the video for a song, the software has no way of knowing that. It’s merely intended as a supplement to the way the music industry already does business.

The company says that Hit Song Science cannot create music of its own. I believe this, but with a few modifications, I’m sure it easily could. If the company doesn’t do it, somebody else will. Already, the program is available in recording studios. As songs are created, producers can now run the songs through the application and get instant feedback. “It helps them to calibrate their music to the widest appeal,” Reed said. I asked her if this could stifle creativity in music, and she told me that most producers are having a lot of fun with it. “They find it helps them to be more creative. It helps to eliminate second guessing.”

You know how it is when you’re jamming with your band, and you’re just a few “underlying mathematical patterns” away from greatness. Polyphonic HMI hopes to help. You want to be a pop star in the 21st century? Better start studying your calculus.

About the author: Joe Lavin writes a weekly humor column at www.joelavin.com. This article originally appeared in Computoredge Magazine.