Hello. In my last blog, I began my discussion of Pandora.com, the streaming audio website which offers a new kind of web radio to listeners. Enter a “seed” song into Pandora’s search engine, and the site will create a streaming “station” composed of songs that resemble your seed song. This process is powered by the Music Genome Project, a massive research endeavor which began in the early 2000s and is based out of the company’s Oakland, California headquarters.
How is Pandora’s song-recommendation engine different than web radio platforms that came before it? Well, the majority of other online radio stations, such as last.fm, operate off a system called collaborative filtering. What is collaborative filtering? In layperson’s terms, collaborative filtering involves matching one user’s taste to another’s (or a series of other people). On a site like last.fm, over time a user amasses a playlist of songs they’ve expressed a preference for—a sort of musical taste profile. Last.fm’s search tools automatically identify other users with whom your tastes seem to overlap, and uses this information to power “radio” stations you can stream on the site. The process is pretty simple, and based on personal intuition and the data existing users have already entered into the system. Collaborative filtering powers aspects of many media websites, such as Amazon.com’s personal recommendation feature for shoppers.
It does have some limitations for online radio listeners, however. As Pandora’s founder Tim Westergren pointed out in a 2006 interview with Leo Laporte and Amber MacArthur, collaborative filtering-powered online radio stations have a tendency to only recommend what is broadly popular in contemporary pop music. While independent-label music certainly has a strong presence on last.fm, a quick scan of various users’ profiles on the site may suggest that Westergren has a point. Even among “indie” users on last.fm, there’s a whole lot of Death Cab for Cutie and Modest Mouse ruling the playlists (nothing against either of these bands). Collaborative filtering doesn’t necessarily ensure that the site’s users will discover truly obscure stuff they hadn’t heard of before. And in keeping with my interest in genre boundaries vis-à-vis Internet radio, in interviews Westergren has attributed the problem to the mainstream music business’ interest in keeping consumers bracketed into genre-specific niches. In the aforementioned chat with Laporte and MacArthur, Westergren cited the “age-old problem in the music industry” wherein a tiny percentage of music released by a given label typically accounts for nearly all its sales—a problem codified by genre boundaries.
Pandora, through its Music Genome Project, aims to circumvent this problem, by offering its users a new kind of recommendation engine. As I mentioned in my earlier post, the Music Genome is a systematic endeavor to deconstruct and analyze individual pop songs using over 400 “musical attributes” that the company has identified. These attributes include everything from tempo, to vocal timbre, to harmonic movement—even sound production aspects like echo and reverb. In other words, it is essentially a musicological approach in the strictest sense of the word. The focus is on sound itself, rather than a band’s cultural associations with other bands (as is the case in collaborative filtering). Indeed, Westergren bragged in the aforementioned interview that “when we recommend to you a piece of music, we don’t even know how popular it is.”
Instead, what the Music Genome Project entails is the company’s roughly fifty analysts sitting down in the Oakland, CA headquarters and methodically tagging a given song using these 400+ attributes. Westergren has described the process in ways akin to the scientific method, noting that a percentage of songs the analysts deconstruct are reviewed twice for quality. The songs, categorized by attributes, are added to the Project’s over 500,000 songs (and counting) accumulating in the company’s database. Songs sharing a similar musical “DNA” are then automatically matched and linked by Pandora’s search engine when you enter in a “seed” song. Westergren has called the Genome “kind of like a musical taxonomy,” and I don’t think this language is accidental. As Fabian Holt has pointed out about musical genres, “Discourse on the temporal dimensions of categories is saturated with organicist metaphors, as in discussions of how genres are born, how they grow, mature, branch off, explode, and die.”1 Even though Pandora in fact aims to get around genre, it seems to me that this biologic language informs the company’s mission and direction.
In any case, as a Pandora user, I have often benefited from the happy accidents occasioned by the way the Music Genome Project works. For instance, I entered in Pandora as a “seed” Bob Dylan’s song “Tonight I’ll Be Staying Here With You”, a lilting, mid-tempo country-rock stroll. The Genome built a streaming station for me that included folk-rocky chestnuts by relatively obscure ‘60s and ‘70s groups like UFO and Earth Opera. It’s likely that I would not have heard about these groups without Pandora, or at least that I would’ve heard about them years from now in another context. In this regard, it seems that Westergren does have something to boast about regarding his claim that Pandora’s search engine connects listeners with “invisible” music in a way that mainstream, genre-bound, multinational music corporations just can’t.
On the other hand, there are several notable gaps in the logic and execution of Pandora and the Music Genome Project model. The first gap I feel compelled to point out is a very practical one. Returning to my example of the station based around “Tonight I’ll Be Staying Here With You,” Pandora is skilled in giving a user a lot of what they like. Enter in a twangy rock song like Dylan’s and you’ll get a station with loads of twangy rock songs. But there can be too much of a good thing; namely, I find homogeneity of songs’ tempo an issue on Pandora stations. “Tonight I’ll Be Staying Here With You” is a bit plodding, and I’ve found that over a few hours of playing this station, I mostly get one plodding song after the next. This can be useful in terms of finding hidden gems, but makes for monotonous, even frustrating listening over a span of a few hours.
I do know that Pandora makes much of its “Thumbs Up/Thumbs Down” feature, which allows the user to indicate her or his preference for a given song. Pandora’s algorithms will adjust the playlist’s direction (ever so slightly) upon a “Thumbs Down” for a song you don’t care for. In a 2006 interview with the New York Times, Westergren describes this feature as a concession to human subjectivity (within an otherwise “objective” platform), and I agree. The “Thumbs Up/Thumbs Down” feature requires active listening and participation on the user’s part—generally a good thing, I’ll admit. But what if I want to just sit back with a cold beverage and let the music play? The Genome’s platform, as it currently works, seems unable to deliver the ebbs and flows in tempo and musical texture which I enjoy in a good mixtape or college radio show.
These kind of practical gaps in Pandora’s service point me toward a larger theoretical problem worth discussing. In its insistence upon musical sound as the key ingredient for making song recommendations, the Music Genome Project willingly suspends belief in some basic social facts about the way music works. Music is undeniably social, cultural, and political. It’s the soundtrack to our lives as we dance, eat dinner, exercise, commute to work, fall in love, and so on. Music blasts out of loudspeakers at political rallies. We argue with friends over drinks about the relative merit of this or that musical group. And music is always part of a commercial marketplace, even in this age of file-sharing. Given all this, I find Westergren’s claim that a band’s marketplace popularity is “completely irrelevant to what we do” a wee bit disingenuous, or at the very least requiring a willing suspension of disbelief regarding music’s social and marketplace role.
The Music Genome Project’s near-exclusive focus on sound itself, coupled with its organicist rhetoric regarding “musical DNA”, seems to suggest the company believes it can map out music in its totality—that it can “crack the code” of music, so to speak. As an aspiring musicologist, this reminds me a bit of another massive scholarly endeavor which worked toward a similar goal of cataloging music: Alan Lomax’s Cantometrics project. Developed by Lomax in the late 1950s and into the ‘60s, Cantometrics was a project wherein Lomax and several co-researchers analyzed the performance styles of (mostly traditional “folk”) songs from hundreds of different cultures around the world, tagging them with a variety of traits. These performance traits, such as vocal timbre, were organized into a computerized system wherein elements of the different musics could be compared. Lomax made the bold claim that one could draw conclusions about the social structure of a given society based on some of these performance traits (societies noted for a certain style of singing were sexually repressive, for instance). Of course, this claim was quite controversial, and has been challenged by other scholars since as overly reductive and essentialist. Fortunately, the Music Genome Project doesn’t attempt to make the connection between music and social structures the way Cantometrics did; indeed, as I said, the Genome Project’s rhetoric seems to deny aspects of the social world, if anything.
However, in the desire to systematically categorize and compare different aspects of music, one could say that the Genome Project and Cantometrics spring from a shared wellspring of human curiosity. One issue with this categorizing mission, though, is the problem of sample size. Lomax’s research was criticized for not casting a broad enough net in collecting these comparable performance traits. One could ask similar questions about the Music Genome Project’s scope. As I mentioned, the company’s website points out that its database currently features over 500,000 songs, and counting. This seems like a lot, but how useful is that number, when one considers the thousands and thousands of songs which are released commercially every year? And how can one ensure that multiple varieties, styles, and (yes, even) genres of music are adequately represented within those 500,000 songs?
Additionally, Pandora shares potentially problematic assumptions with Cantometrics regarding humans’ ability to fully categorize and catalog the world, to reduce music to its essence. This descends from the Enlightenment idea that our natural world is fully knowable through empirical and objective observation. That bedrock assumption has been the basis for the natural sciences, and one can see its influence on a project like Alan Lomax’s. The problem is: while an empirical observation approach might work well for classifying different varieties of tree frogs, when one wades into the murky waters of human behavior, it’s a lot more difficult to claim objectivity. Indeed, for myself and for a growing number of musicologists and humanities scholars more broadly, it is basically impossible to claim objectivity in one’s understanding of the world.
This is not to say that Pandora explicitly makes a claim of total objectivity, on their website or elsewhere. But as with Cantometrics, the fact that the company breaks songs down into discrete components and then makes comparisons and connections based on those components suggests that they believe music is knowable in some objective way. Pandora hasn’t made public a list of its over 400 “musical attributes”, but shares a handful of them on their website’s Frequently Asked Questions page. Some of the attributes they share make a lot of sense, and could even be called “objective”: major or minor key tonality, for instance. But consider an attribute like “headnodic beats”: in its FAQ entry, Pandora’s analysts admit they created the term themselves (it describes hip-hop beats which are strong, but not forceful enough to dance to). Given that probably almost no one outside of the Pandora offices uses this term, it can’t reasonably be called objective. This is not to say that an identification of some subjectivity within Pandora’s research model makes the whole enterprise come crashing down. Rather, I just wish to point out that while company prides itself on the cold objectivity of a computer algorithm choosing your music for you, human beings with subjective viewpoints created the components which power that algorithm.
Related to this, both Pandora and Cantometrics raise questions regarding musical gatekeepers, tastemakers, and their authority. Indeed, as ethnomusicologist Steven Feld has mused in response to Lomax’s work, “What are the sources of authority, wisdom, and legitimacy about sounds and music? Who can know about sound? Is musical knowledge public, private, ritual, esoteric?”.2 Many researchers of pop music, pop culture, and genre agree that this issue of who is doing the classifying, categorizing, and ranking is a really important question. For instance, snobby clerks at your local independent record store may decide that Gillian Welch’s music belongs in the “folk” rather than “rock” section of the store. But where does the authority behind their judgment come from? Their judgment is informed by their life experiences and backgrounds as (mostly) well-educated middle-class white males.
These gatekeepers’ judgments are also informed by a deep knowledge of various musical genres: the ability to distinguish glam from punk from grunge, and so on. Since Pandora’s fifty music analysts perform a similar function, I find this aspect of their job paradoxical: though the company seems to pride itself on getting beyond musical genre, these analysts must be extremely well-versed in genre in order to do their jobs well. In a recent video post on Pandora’s blog, Westergren states that the purpose of the Music Genome Project is ultimately to connect musicians with audiences, in ways the traditional music business can’t.3 This is notably egalitarian rhetoric; it works off the assumption that consumers and musicians are empowered enough to seek each other out, and that they don’t need tastemakers dictating what music they should like.
As I noted above, however, the computerized system that listeners use to connect with musicians is designed and maintained by a group of (relatively) elite tastemakers. And in Westergren’s public statements about these analysts’ qualifications, I read a certain degree of anxiety over what kind of authority is vested in that role of analyst. In his 2006 interview with Leo Laporte and Amber MacArthur, Westergren pointed out that while all their analysts are regularly-gigging musicians, in order to carry out the depth of analysis required for the Music Genome Project, one really “need[s] an academic background”. Thus, in addition to being a working musician, an analyst employed by Pandora also needs at least a four-year undergraduate degree in music theory.
On a practical level, this makes a lot of sense to me. If you’re going to employ folks to analyze songs for you, wouldn’t you want them to have an understanding of musical principles on several different levels? On the other hand, on a theoretical level, Pandora’s insistence on both “street” and “book” smarts from its analysts demonstrates an unresolved subliminal conflict over whether “brains and corporate no-how” or “gut, ‘Id’ feelings” are what shape the music we listen to. Thus, in this way, Pandora and the Music Genome Project struggle with these issues of taste and knowledge hierarchies just like other public pop prognosticators, even as their seemingly objective research platform denies this social fact.
This may read as though I am beating up on Pandora, but I hope the position I’m staking out is subtler than that. Rather, I have simply been attempting to point out some slight contradictions of logic within the Music Genome Project’s overall research platform. On a practical, user’s level, I enjoy the site. And to be fair to Pandora’s employees, on a certain level they seem to recognize the issues I am brining up here. For instance, in a recent post on Pandora’s official blog by one of its music analysts, Michael Zapruder likens evaluating songs to judging a baby beauty contest, and then points out,
The idea that all music is equal and deserves equal rights is somehow fundamentally a democratic idea; as is the corresponding idea that the public, and not some small cadre of experts, is the best judge of musical quality. But the fact that some music not only attracts more listeners, but also seems to mean more to more people over a longer period of time, indicates that there is actually something fundamentally unequal about music as well.4
In other words, perhaps this issue of taste isn’t an “either/or” problem, but rather a “both/and” one. And by its nature, it’s most likely a problem with no definitive answer.
It seems that in his blog entry, Pandora employees like Zapruder are trying to find a practical, everyday way of working around and through this problem—and I can’t fault them for that. Certainly, the academic in me bristles when I see Pandora present something like “headnodic beats” as some kind of objective criteria for judging music. But on a practical level, it seems that these classifications, even if they’re vague (such as “vinyl ambience,” or what have you) are perhaps vague at least partly in the service of the listener’s experience—of trying to match users to interesting new music. It doesn’t seem that the point of the Genome is to categorize musical attributes simply for the sake of categorization. Rather, the point seems to be to put that information to use, making musical connections for the listener. So perhaps it’s a utilitarian reason why the Music Genome cuts certain logical corners on the “objective vs. subjective” question.
Ultimately, Pandora’s service rests upon the assumption that sound itself is the only aspect which really matters when analyzing different forms of music. It assumes that sound automatically trumps the sociocultural boundaries of genre, taste, and marketplace. This isn’t true, of course: in the real world we live in, rhetoric surrounding genre and taste guide the musical choices we all make, from Walmart AC/DC lovers to bebop nerds. But the Music Genome Project’s fiction regarding the supremacy of sound is an important, if very one-sided, position to have out there in the world. In fact, it’s almost counter-cultural in a way, because journalists and advertisers often focus so much on image when considering contemporary pop music. Pandora’s vision is a kind of imagined musical utopia, making a particularly 21st-century-specific stand for the importance of musical sound—a stand made possible by the shared cultural resource of the Internet.
Finally, closing with an idea my professor Fred Maus pointed out to me, when you’re confronted with enjoying a song you didn’t think you would like on Pandora (you enter in a Turbonegro song as your “seed” and are rewarded with a Poison song, for example), that tells us something important about genre boundaries. Your bemusement proves that musical genres exist. They’re cultural; they don’t hold up to objective scrutiny. And they’re based on something more than just musical sound; they’re built around assumptions that have to do with hierarchies of taste and class. Thus, paradoxically, we can learn quite a bit about the rules of genre from a website devoted to transcending those rules.
Holt, Fabian. Genre in Popular Music. Chicago: University of Chicago Press, 2007. Pg. 14. ↩
Feld, Steven. “Sound Structure as Social Structure.” Ethnomusicology, Vol. 28. No. 3 (Sept. 1984), pp. 383-409. ↩
http://blog.pandora.com/pandora/archives/2009/03/index.html March 15, 2009 entry. ↩
http://blog.pandora.com/pandora/archives/2009/02/index.html February 25, 2009 entry. ↩