Technology is continually getting better at predicting our behavior, and signaling to us that perhaps we’re not as unpredictable and free-willed as we humans like to think.
Software can predict where crimes are going to happen, scanners can guess what products we like based on what we look like, and brain scanners can eavesdrop on our inner dialogue just by reading our brain activity. One could plausibly argue that the aspect of human nature least likely to be called formulaic is artistic creation, such as making music. Researchers at the University of Bristol, however, want to turn our inspired musical notes into cold, calculated plusses and minuses. They’ve come up with an equation that predicts whether or not a song will become a hit. True reductionists that they are, the scientists plan to use the equation, not only for prediction, but eventually for production.
The equation, dubbed the “Hit Potential Equation,” seeks to break down songs into their component parts and then identify the components most important to hit potential. Parameters such as loudness, tempo, beat variation and “danceability” are assessed for their power to predict if a song will be a hit – reach top 5 on the hit charts – or become a flop – reach no higher than #30 on a top 40 list. The equation uses 23 distinct audio features in all to score a song’s hit potential.
How accurate is the equation? It correctly predicted when a song would make it into the top 5, or not reach above 30, on the UK Top 40 Single Chart 60 percent of the time.
Seems our musical leanings are too complicated to be captured in a single equation, at least for the time being. It’s not surprising that song popularity is not determined solely by their audio features. Like the comeback of the exposed midriff, style trends are affected by a multitude of factors, not just the number of sequins on a dress or the harmonic simplicity of a song. You’re more likely to enjoy country music if you grew up in Nashville rather than Boston. A song’s popularity also affects, well, its popularity. A 2006 study looked at how the appeal of a new song to a person was affected by how much others liked, or didn’t like, the song. They showed that whether or not study participants were aware of previous participants’ choices made a huge difference. Social influence led to the unpredictability of a song’s success.
And as we all know musical tastes change over time, which means the Hit Potential Equation has to change over the years if it’s going to have a chance at uncovering the next Lady Gaga. The audio features that were useful in picking the top 5 in 1961 are quite different from those useful in 2011. The scientists produced a video showing the shifting predictive powers of the audio features from 1961 to 2011. Unlike our parents, we like it loud.
The work was led by Tijl De Bie (pronounced “Tell De Bee”), a senior lecturer in artificial intelligence at the University of Bristol. Part of the Pattern Analysis and Intelligent Systems Laboratory, Bie is interested in using AI to recognize patterns in large bodies of data and then using that information. At the root of the Hit Potential Equation is a branch of AI called machine learning. The computer takes the hottest songs and “learns” what audio features they have in common.
Their analysis turned up some interesting musical trends over the years. Danceability, never before important in discerning between hits and flops, became important in the late 1970s when disco had its mercifully short period of popularity. Around 1980 the equation performed terribly, dipping to a 52 percent low. They speculate that the change in musical style from the late seventies to the early eighties was particularly pronounced, that the time was marked by exceptional innovation and creativity of pop music. The equation redeemed itself in the early ‘90s and year 2000 when it performed the best, coinciding with a decrease harmonic complexity. Pink Floyd, harmonically complex. Nirvanna, not so much.
Tracking loudness it’s pretty clear that music has been getting steadily louder since the 1960s. Right about 2008, though, the graph turns downward. Perhaps there’s a hidden ‘neighbor factor’ that’s causing us to not crank it so loud.
Taking a look at specific songs, the equation correctly predicted “Crazy” by Gnarls Barkley, due to its danceability, low energy and loud signal. “Crazy” was a number 1 hit in 2006 for 6 weeks. The researchers classified “Crazy” as an “Expected Hit,” a song that matches a hit profile so well that they don’t need a computer to know it’s going to be a hit.
Other songs the equation gets completely wrong. It predicts a flop but in reality, the song indeed climbed the charts to the top 5. These “Unexpected Hits,” they theorize, do well because of factors that their equation doesn’t consider such as social context. One example was the 2009 release of Michael Jackson’s “Man In The Mirror.” It did modestly well when it was originally released in 1988, reaching only as high as #21. Following Jackson’s death, however, the song peaked at #2. They say that the song should have been “too quiet and insufficiently danceable” to reach such a high mark.
Like any good scientist, Bie and his colleagues are loyal to their equation. When it predicts that a song should have appeared in the top 5 yet history shows that the song be a flop, they argue that listeners just got it wrong. They call these songs “Hidden Gems” that could be rereleased because of their hit potential. An example they give is “The First Cut Is The Deepest” by Sheryl Crow. I bet you didn’t know it was originally written by Cat Stevens in 1967. But cover versions of the song by P.P. Arnold, Keith Hampshire, Rod Stewart, and finally Crow in 2003 were all more popular than the original. They argue that the fact it became a popular cover several times over shows that the original had a potential that was unrealized at the time, a potential pointed out by their equation.
Bie and his colleagues want to share in the musical data mining fun. They’ve built an app, called scoreahit, with which anyone can score songs in their own collection or even songs they produce if it’s in mp3 format.
Because musical tastes change over time, predicting a musical hit is like hitting a moving target. To keep their equation as accurate as possible it is constantly recalibrated based on the success of recent releases. Additionally, the lab is trying to incorporate emotional content of lyrics into the equation. They think whether or not the song is uplifting or wallowing in its own sorrow will affect its appeal.
Despite the fact that the scientists are simply using music as a data source with which to test their data mining and AI capabilities, Bie thinks the equation can be a benefit to music makers. Don’t worry, he doesn’t envision a future in which computers replace inspiration as the generators of musical templates. But he does hope that scoreahit can help smalltime garage bands gauge whether or not their latest creation has hit potential. He hopes it can help them break into the market. The more people making music, he argues, the better for music as a whole.