Congratulations to Lil Nas X: “Old Town Road” has just been voted as the No. 1 “Song of the Summer” according to Billboard. It’s taken a long journey from unassuming SoundCloud beat to Billy Ray Cyrus remix to Ellen DeGeneres meme, but “Old Town Road” now appears to officially be at its zenith as a worldwide hit.
But is it a country hit? Nay, declares the Billboard Hot Country chart. Back in April, “Old Town Road” got booted from the list for not embracing “enough elements of today’s country music to chart in its current version,” per a Billboard statement. The hot takes poured forth on all corners of the culturesphere, from Reddit to national newspapers to YouTube to Rolling Stone. But thank goodness: Now there’s an objective, rigorously scientific source to clear the air for us.
According to a new computational tool, “Old Town Road” is — trap drumroll please! — pop music. Pop music with totally country lyrics and an unmistakably rock-and-roll guitar part, but pop music nonetheless. In fact, it’s those diverse attributes that make the tune a perfect medium to examine how humans categorize the sounds and emotion of music in our never-ending mission to organize and label the things we love.
This new A.I. musical overlord exists courtesy of Timothy Greer, a computer science Ph.D. student at the University of Southern California, who built musical genre-prediction software and then “trained” it by making it process a dataset of 190,165 musical sections from 5,304 songs with lyrics and corresponding chords. There are three computational models used in this experiment: One that reviews only lyrics; another that reviews just music; and a third that reviews the interplay of the two. It’s the latter that proved most revealing, Greer says.
It’s all possible through a technique called “machine learning,” which basically automates data analysis with minimal human intervention. Genre-prediction algorithms aren’t new, and neither is scientific research on how music affects emotion and cognition. But Greer and his supervisor Shri Narayanan, director of USC’s Signal Analysis and Interpretation Laboratory, have produced a project that searches for links and patterns in lyrical content and instrumentation, together — potentially uncovering a path to discover how the human brain can so quickly calculate something as cognitively complex as genre.
“Basically, if you have an A-minor chord, the lyrics you would choose to write over that may be way different from what you write over, say, E-major. That sort of information is used in this study to evaluate musical perception,” Greer explains to me over Skype. “I think that’s a real breakthrough in research in this field because it’s about the relationship of two languages, if you will, that move separately.”
The idea came from Greer’s project from earlier this year, which used a similar analysis of lyrics and instrumentation to predict listener emotions for any given musical passage. It also builds on other research that Narayanan and Greer have conducted together, which includes doing MRI scans of beatboxers and building a program that can identify humor.
There are limits to how practical such tools are today, but the use of machine learning helps us see hidden wrinkles and ask new questions, Narayanan says. “Our lab is leading research on computational media intelligence, trying to understand the stories we tell in different formats. We love to label genres in music because humans try to categorize things innately. We love patterns,” he explains. “Figuring out whether ‘Old Town Road’ is hip hop, pop or country… Well, they share common characteristics and a song never neatly falls into one category. That’s what Tim was able to show, computationally.”
There are obvious results and brow-raising surprises in the way Greer’s model predicted genre and related songs to one another, all displayed on a scatter plot (go here and scroll down to click around the chart yourself). It’s difficult to understand why Tom Petty’s “Runnin’ Down a Dream” is a rock song, while Green Day’s “Boulevard of Broken Dreams” is a pop track. Yet the closeness between those two particular dots indicates they have a lot of things in common in the music and lyrics; indeed, you can hear the resemblance in all that melancholic singing about roads and dreams. It’s more fun when the genre label and associations just feel wrong — as with electro-alt-rock-ish band Fall Out Boy’s “Young and Menace” being categorized as “hip-hop” (?????), nestling right up next to… Florida Georgia Line’s decidedly bro-country “Cruise” (?!?!).
More thoughtful revelations abound: Flo Rida’s ode to blow jobs “Whistle” is a country tune, folks. (This one makes sense, given the jangling acoustic guitar and constant thinly veiled metaphors about gettin’ some.) So is Maroon 5’s “Won’t Go Home Without You,” an aggressively mild but seemingly un-country ballad. “Viva La Vida,” that syrupy Coldplay song you were trying to forget over the last 12 years, is hip-hop. “Believer” by Imagine Dragons, meanwhile, is apparently a rock song? News to me!
It’s important to remember that a genre-prediction tool, no matter how deep its machine learning, is a jumping-off point for more complex research and debate. Greer notes that part of his curiosity about music and lyrics comes from the fact that one can “trick” the other — sad words take on new meaning when paired with bright, optimistic chord voicings rather than somber ones, as he points out. Musicologists have long pondered how music can inspire such vivid emotions and memories despite the medium lacking the capability to depict reality like a film or painting can. Even mere intervals between single notes can have an emotional quality to them — as musicians and YouTubers Adam Neely and Ben Levin investigated in a recent video.
And it’s key to consider that genre itself is a hugely important attribute that’s manipulated by artists, agents and brands all the time. Lil Nas X’s own manager, Danny Kang, more or less proved that point. Lil Nas knew what he was doing by releasing “Old Town Road” as yeehaw meme culture really caught fire, Kang observed to Rolling Stone, implying that labeling the tune as a country track when uploading it was a savvy move to stick out. “On SoundCloud, he listed it as a country record. On iTunes, he listed it as a country record. He was going to these spaces, gaining a little bit of traction on their country charts, and there’s a way to manipulate the algorithm to push your track to the top,” Kang said. “That’s favorable versus trying to go to the rap format to compete with the most popular songs in the world.”
The perception of genre in the human brain depends on what you’ve heard, what you know you’ve heard, and whether you pay attention to details like, well, lyrics. To my ears, “Old Town Road” is a pop-trap tune with folksy old country lyrical stylings, but finding the “right” answer is maybe missing the point. Our arguments over genre, whether in music, film or literature, is the result of the sheer diversity in how we consume and store art in the attics of our minds. Greer thinks his computational model could one day be useful as a tool for songwriters, or for the generation of smart and surprising playlists. Narayanan takes a bigger view: “Music is very innate to us humans, spanning cultures and religions, so we’re very interested in how we create, perform and enjoy it,” he tells me. “From a scientific perspective, machine learning gives us the tools to find that understanding.”
Still, I have to wonder what the model would say about this properly countrified cover of “Old Town Road.”