FEATURE CREEPS
How Pandora’s Music Genome Project Misrepresents the Way We Hear Music

When you’re a Silicon Valley tool with a hammer, everything looks like a nail. Among other things, this means that nearly any object or phenomenon known to man is reducible to some formula or algorithm, and thus replicable—only better!—in a lab or startup garage. This geek chutzpah brings us things like “Food 2.0,” a movement toward creating synthetic replacements for much of the world’s food that was touted in the Times recently and has attracted some of the world’s wealthiest tech investors. Undeterred by the long history of can’t-miss science experiments that have ended up having disastrous consequences for the environment or people’s health, we have the plucky, all-knowing chief executive of the San Francisco-based food-engineering company Hampton Creek, who modestly boasts, “We’re looking for wholesale reinvention of this crazy, perverse food system.” And who better to correct two billion years of barely understood evolution than a budding startup mogul?

Another grand-scale techie dream that thankfully doesn’t threaten the food chain, but is only slightly less frightening for people who care about music, is embodied in Pandora Radio’s Music Genome Project (M.G.P). The mission of the M.G.P. (first conceived around the turn of the millennium) is to isolate and catalog the significant characteristics of every piece of music in Pandora’s online library, with the goal of identifying the features that appeal to individual listeners and using that information to make new recommendations. If it’s determined that you tend to like, say, songs in a minor key, or ones with long instrumental passages, or those that feature electric-piano solos, Pandora’s engine can easily use its database to find more songs that tick those boxes and recommend them to you. Of course this is only one of the many approaches to online music “discovery” that are being tested by music services, and it remains to be seen how well it will work. It may turn out that the much lower-tech method of simply telling you what tracks your online friends are listening to will be more effective. But that rather large issue aside, the whole project, and the idea of this kind of feature-analysis of music generally, raises some interesting questions, and some disturbing ones.

First and foremost, it seems clear that Pandora’s attempt at a complete music taxonomy, beyond the grossest kinds of generalizations, is literally impossible to pull off. While the actual list of song-features that Pandora keeps tabs on is a company secret, it’s very long—a 2010 article in Fast Company already lists 400-plus attributes. And the analyses are performed by people, most of them seasoned experts rather than machines (which would really be impossible). But what can these analyses consist of? They certainly include the basic qualities mentioned above: major-key vs. minor-key, vocal vs. instrumental. And the Fast Company article shows an example of some deeper digging by one of Pandora’s resident musicologists—a harmonic analysis of Herbie Hancock’s “Maiden Voyage”: “The song is built entirely with suspended dominant chords, which have an ambiguously minor sound. The lack of resolution gives a floaty quality to the song.” But however many of these characteristics Pandora itemizes, can they ever be enough? This is not an esoteric question.

The kind of musicological analysis (harmonic, structural, etc.) described above is well established and perfectly reasonable as far as it goes, but the Music Genome Project is not purely about static descriptions of how a piece of music is constructed—or at least it shouldn’t be. If the goal of the project is cataloguing the attributes that might possibly lead a listener to want to hear a song over and over (leaving aside, again, the important question of whether this is how people’s preferences are actually determined), those attributes can be exceedingly difficult for a handful of analysts to isolate, and in fact the list is effectively infinite.

For example, a listener’s pleasure center may be activated by a subtle, or barely perceptible, vocal inflection, or a style of guitar strumming, or a kind of amplifier distortion. Furthermore, any of these innumerable quirks may constitute a kind of secret handshake (or signifier) among some cultural clique. There are countless precedents for this, especially in rock and pop music: Guitar solos get longer, then shorter, then disappear, then reappear. Melisma becomes a non-negotiable requirement for female singers, then singing becomes (must be) completely affectless. Record production lards on reverb over the years, then cuts it away, or switches to a drastically different convention for creating reverb (compare the overall sound of an LP recorded in one of the huge studios of the 1950s to a digital recording made recently). There are sub-genres, and sub-sub-genres, and inchoate genres, where even the particular, nuanced way of recording a snare drum creates an unbridgeable chasm between the way music sounds today and the way the dinosaur groups of a couple of years ago sounded. Not only are the rules that define what features matter often unknown outside some specific social group—imagine a 70-year-old Republican resident of Orange County attempting complete descriptions of five hip-hop songs—but awareness of them is often subconscious within that group. No doubt many people tend to like songs that are “sad,” or short, or have repetitive choruses, but those features, even if they number in the hundreds, are just the tip of a very large perceptual iceberg.

Cultural myopia can also play a significant role in the ostensibly objective act of defining a song’s distinguishing features. One analogy can be found in the branch of linguistics called phonology. Phonetic features that have tremendous significance in one language often have none at all in another. In Cantonese, for example, there are multiple tones (roughly, inflections), and switching from one to another in pronunciation, while leaving everything else the same, changes a word to a completely different one. There is no equivalent of this in English—changing the tone cannot change the word—and as a result people raised as English speakers can be unable to detect differences in tone in Cantonese. On the other hand, the classic confusion between l and r among many Chinese people reflects a distinction in English that isn’t present in their native languages, and as a result has become impossible for them to hear. There’s nothing racial about this: The sensory input we’re confronted with all day long is ridiculously complex, and for the sake of sanity people learn from an early age which pieces of information are signal and which are noise in their culture. But making assumptions about the set of distinguishing features of a piece of music based on what is necessarily limited knowledge and cultural experience will always mean dropping many significant scraps on the laboratory floor, like a model car with a dozen parts left out. In his book Improvisation, the great English guitarist Derek Bailey discusses the difficulty Western musicologists have had in analyzing Indian improvised music, and attributes this partly to the concept of laya in raga, which Bailey describes as commonly thought of as tempo of a piece, but is “much more than that. It is its rhythmic impetus, its pulse. The musician who displays an exceptional rhythmic ‘feel,’ whose work has great rhythmic facility and ease, is described as ‘having a good laya.’ The vocabulary of Western classical music contains no equivalent for laya, either being incapable of recognizing its existence or preferring to ignore it.” This kind of cultural blindness, Bailey claims, has frequently led to incomplete descriptions when these academics have attempted what they considered an authoritative analysis of this music.

Getting back to changes over time within Western music, features that mark two pieces as drastically different today may fade to insignificance in 10 years. Think of a broad genre like 1940s swing, or 1960s – ’70s Classic Rock: During those periods even casual music fans would have noted clear distinctions between different groups’ music, and less casual fans may have literally duked it out in the streets over rival bands. Late-’70s radio stations specializing in Album-Oriented Rock might have been deluged with outraged letters if they’d played something that was a little too punk or (heaven forbid) disco, as defined by the sound of the hi-hat or bass guitar. But today the similarities between these once-warring camps are often seen as much more important than the differences to anyone but the most obsessive music historians, and maybe even to them. (For a hilarious illustration of this, listen to Andy Partridge’s 20-second-long “History of Rock and Roll,” which distills 40 years of rock history into four representative sound-bites.) In the late 1950s, Miles Davis probably spoke for many jazz fans when he said of the avant-garde music of upstart Ornette Coleman, “If you’re talking psychologically, the man is all screwed up inside,” but in 2014 it seems unlikely that many Miles fans would take much notice of the differences that were irreconcilable to him. So a Pandora analyst creating a “genome” of an Ornette track in 1959, or a big-band record in the 1940s, or a Sex Pistols song in 1976, would certainly come up with a different list of significant features than he or she would today—and rightly so, because different features would have been meaningful, or even detectable, to fans of the time. Music-listeners don’t exist in a cultural or temporal vacuum.

If these obstacles to creating a full set of music genomes really are insurmountable, and Pandora’s approach to music discovery is destined to merely hobble along as a result, one inevitable question is, Who cares? This isn’t Silicon Valley’s first over-ambitious, data-obsessed business plan, and it won’t be the last. Does it make a difference to anyone without a financial interest in Pandora? Well, it might. Even if the Music Genome Project ends up not being useful in music discovery, its very 2010s Big Data approach to music will have its appeal to many other tech-worshipers. (Netflix, interestingly, is trying something similar—but more benign, in my opinion—to tag its movies.) It’s a good bet that would-be hit-makers looking for the magic formula for that elusive #1 record are paying attention, thinking of ways to apply Pandora’s model by creating songs that contain only the important features, as determined by some M.G.P.-like list—the model car with a dozen missing pieces again. But more generally, despite all the high-blown talk about algorithms, this model constitutes a sadly reductionist, anti-holistic view of what music is, like any number of previous attempts by self-proclaimed savants to remove nature’s messiness, generally with profit in mind. Will M.G.P.-style descriptions of music be to the 21st century what the miracle of Wonder Bread was to the 20th?