Quote:

Each song in the Music Genome Project is analyzed using up to 450 distinct musical characteristics by a trained music analyst. These attributes capture not only the musical identity of a song, but also the many significant qualities that are relevant to understanding the musical preferences of listeners… The Music Genome Project’s database is built using a methodology that includes the use of precisely defined terminology, a consistent frame of reference, redundant analysis, and ongoing quality control to ensure that data integrity remains reliably high. Pandora does not use machine-listening or other forms of automated data extraction (About the Music Genome Project).

Pandora Radio is not your ordinary internet radio. Like many others I jumped on the Songza internet radio train for a long while and believed it to be the ideal solution for throwing on a suitable playlist at a moment’s notice. But then (regrettably) just a couple of weeks ago one of my(?) students showed me the error of my ways and recommended I check out Pandora Radio instead, which has indeed been around for some time now. Unlike Songza which primarily prompts its users to choose playlists based on what activity they are presently doing (e.g. Entertaining cool friends, unwinding, recovering from last night etc.), Pandora prompts its users to enter in an artist, a genre of music, or even just a single song, and it then creates a radio station based on this single piece of information, a station that plays songs that are of a similar nature to said genre, artist, song etc. That’s right, your favorite song can now serve as the foundation of a playlist. If that’s not enough for you, you can then click the “add variety” button, type in another artist, song etc. and now your Led Zeppelin playlist can have a touch of Frédérik Chopin. Don’t like a song that Pandora offers you? Just click the thumbs down button and Pandora will keep playing your custom playlist but will now avoid songs that match the parameters of the song you disliked. The opposite also holds true – hit the thumbs up on a song and Pandora refines your preferences further.

Click to enlarge!

Like many others I was amazed the first time I experienced Pandora, so much so that I felt the need to declare on Facebook that Songza should proceed to go ahead and eat its heart out. My first request of a “Death Cab For Cutie” (arguably my favorite band) playlist yielded some bands I already enjoy (Stars, Coldplay, Keane, Iron and Wine), but also Pandora offered me songs that I had never heard by said bands, songs I enjoyed upon my very first listen. But I was really sold when I decided to “add variety” and entered in the band “Every Time I Die” (another one of my top five) as a modifying artist to my Death Cab playlist, an action which prompted Pandora to play a great song from another one of my preferred (sadly, now disbanded) artists, Underoath. A bit of Maylene and the Sons of Disaster along with some other interesting “hardcore” bands I had never heard before followed, and I was done with Songza forever. Now, this pleasurable first experience potentially speaks to some sort of narrowness of musical taste on my part – that I am partial to a select few of these 450 classification parameters – and perhaps this is the case with many of us.

Indeed, I am not the only one to have had this positive first experience with Pandora. Pandora has been steadily increasing in popularity since its online launch in 2005 and reached over 150 million users in 2012. Journalist Julie Layton writes, “With all of Pandora’s attractive features and novel approaches to personalized radio, it does tend to wow people when they discover it.” She also writes of the merit of the user feedback system: “The idea is to continually provide feedback so the station learns more and more about what we like and don’t like. The result is a progressively personalized radio station that really does play only music we want to hear. It takes a while to get there, but most people agree that the feedback process works.” Pandora even lets its users inquire as to why it chose to play certain songs. Just click on the album art, click on the up arrow, and inquire. Apparently, since I decided I wanted a Death Cab For Cutie playlist, Pandora thought I would appreciate “Your ex-lover is dead” by Stars because “it features electric rock instrumentation, a subtle use of vocal harmony, mild rhythmic syncopation, intricate melodic phrasing and extensive vamping,” while it also thought that I would enjoy “The Sound of Settling” by Death Cab For Cutie (well, no kidding) because it “features electric rock instrumentation, major key tonality, a vocal-centric aesthetic and many other similarities identified in the Music Genome Project.” In fact, because of such a mass of classification parameters, Layton recommends that Pandora users create playlists based on individual songs rather than artists (there’s always a few songs we do not like, even with our favorite artists, right?) to more quickly allow Pandora to figure out our musical tastes.

In essence, Pandora (which, for the sake of this probe’s simplicity, I am sometimes conflating with the Music Genome Project… just keep in mind that Pandora is essentially the interface that we use to interact with the database that is the Music Genome Project) is simply a mass digitization of music. It is based on, in Layton’s words, “an intricate analysis by actual humans.” Indeed, according to Pandora’s website, “trained music analysts” who “(have) a four-year degree in music theory, composition or performance, (have) passed through a selective screening process and (have) completed intensive training in the Music Genome’s rigorous and precise methodology” classify each and every song, as aforementioned, according to a collection of over 450 distinct musical traits e.g. arrangement, beats per minute, gender of the vocalist etc. Clearly we are dealing with a digitization process not unlike that which Matthew Jockers describes in his Macroanalysis: Digital Methods and Literary History. Jockers explains the idea of literary macroanalysis and how mass digitization of literature is allowing us to analyze literature in new ways. Jockers explains that since we have managed to digitize such a large body of texts from various centuries, we can now use computers to accomplish literary tasks that were never before conceivable:

Franco Moretti has estimated that of the twenty to thirty thousand English novels published in Britain in the nineteenth century, approximately six thousand are now extant. Assuming that a dedicated scholar could find these novels and read one per day, it would take sixteen and a half years of close reading to get through them all … A computer-based analysis or synthesis of these same materials is not so difficult to imagine (19).

Jockers’ essential point is that we can now have a sort of “birds eye view” on gigantic bodies of literature and so we can, with the right tools, easily uncover elements such as word patterns (e.g. Words that are more likely to pop up in first-person narratives compared to third person narratives), or perhaps trends of genre over time (e.g. The percentage of all novels released each year in the nineteenth century that can be considered fictional {using a set of parameters that predicts the likelihood of a novel being fictional according to its diction, syntax etc.}).

From these efforts, Jockers claims that we may be able to uncover important ideas about some texts that are supposedly “representative” of their time. For example, Jockers speaks about Ian Watt’s research and how Watt argues that “elements leading to the rise of the novel could be detected and teased out of the writings of Defoe, Richardson, and Fielding” (7). Jockers does not reject Watt’s findings and considers them “sound,” but then he speculates about the fallibility of these sorts of claims: “Can we, in good conscience, even believe that Defoe, Richardson and Fielding are representative writers? Watt’s sampling was not random; it was quite the opposite. But perhaps we only need to believe that these three (male) authors are representative of the trend towards “realism” that flourished in the nineteenth century” (8). Here, Jocker hints towards the idea that literary macroanalysis might prove wrong our conceptions about nineteenth-century literature. Perhaps if we could conduct what I described before (a sort of genre test, this time for “realism”) we could potentially discover (this is just a hypothetical example) that what we define as “realism literature” might not in fact have been the main trend of nineteenth-century literature. Perhaps a slightly different genre was actually more widely practiced, and perhaps the writings of Defoe and Richardson became the representative samples of the supposed trend towards realism for some more abstracted reason (perhaps simply because they were elegant writers and so they received the most public praise). In other words, Defoe, Richardson and Fielding being “representative” writers depends in part on there having been a flourishing of “realism” in literature in the nineteenth century, but perhaps there was a more widely practiced genre of literature in the nineteenth century – something we could hypothetically discover with macroanalysis. Whether or not we wish to define “representative works” as those which were the most influential for the emergence of other popular genres, or those which exemplify the genre that writers practiced most in a particular time period, or those which were most well received by the public in their time, or whether or not genre is really what we should be analyzing, is a moot point here. What is important is that macroanalysis has the potential to shatter the conceptions about literature we already have. We should also must keep in mind another moot point, that being the fact that some works are popular in their time whereas others only become popular retroactively and posthumously.

What I find most interesting here are the ways in which we come to value mass digitization. In the digital humanities, we gain the potential for all new information about works written long ago. In music, suddenly we can build and utilize an interface to interact with an intricate database for our listening pleasure. I hope that in the future (unless this has already been done and I am totally unaware) that we can create a Pandora style literature recommending interface, if you will – something a little more elaborate than Amazon’s “Customers who bought this book also bought this book” recommendation method. Perhaps if we can begin to classify literature according to as many parameters as Pandora does music, we can build a program to recommend books based on very specific user input – let us imagine entering in Joyce’s descriptions of hell in his Portrait and getting a list of books we might like based on that query. Oppositely, the Music Genome Project could serve to broaden the field of music study – we could pose questions about music history similar to those that Jockers, with respect to macroanalysis, poses about literature (e.g. whether style is nationally determined, the extent to which subgenres reflect the larger genres of which they are a subset etc.

But there are several problems that we encounter with mass digitization. Ted Underwood points to a big one – that being the fact that digitization involves classification, and classification is “a problem created by us. Distant reading is hard, fundamentally, because human beings don’t agree on a shared set of categories” (par. 13). Layton points out that for all we know, Pandora’s music “experts” are potentially just a couple of blokes “who don’t know the difference between syncopation and vamping.” She also points out that the Music Genome Project is proprietary, and so there is a lack of a more universal review/critique of its mechanics. The question of authority also arises. Who has the authority to classify music as one sort or the other? Who has authority when it comes time to design algorithms that determine how likely a novel is to be, say, fictional? When Underwood writes of the blurriness of categorization, he points to how we can alter algorithms to accommodate different opinions and he uses the example of what constitutes point of view: “Try defining point of view one way, and see what you get. If someone else disagrees, change the definition; you can run the algorithm again overnight” (par. 14). But how many times must we run this algorithm with different initial definitions in order to get “representative” data? Underwood claims that this blurriness is actually beneficial and widens our potential for discovery, a respectable claim I think, but we must keep in mind the subjectivity of each individual as this vastly complicates our distant reading methods. Nevertheless, the first experience of Pandora is, for many, pleasurable – thus it would seem as if this digitization of music has been successful. Have we really accurately mapped the genome of music? Have we really accurately mapped the genome of literature? The answer right now is of course a big NO, but we’re getting there…

P.S. If you want to try out Pandora radio for yourself, you will need an I.P. address disguising program of sorts. I use Hola Better Internet on Google Chrome – you can find it in the Chrome Extension Store for free.

“About the Music Genome Project.” Pandora Radio Website. Web. 29 Apr. 2009.

Jockers, Matthew A. “Revolution,” “Evidence,” “Tradition,” “Macroanalysis.” Macroanalysis: Digital Methods for Broad Outlines of Literary History. Urbana: University of Illinois Press, 2013. 1-32.

Layton, Julia. “How Pandora Radio Works” 23 May 2006. HowStuffWorks.com http://computer.howstuffworks.com/internet/basics/pandora.htm 13 October 2014.

Underwood, Ted. “We Don’t Already Understand the Broad Outlines of Literary History.” The Stone and the Shell. February 8, 2013. http://tedunderwood.com/2013/02/08/we-dont-already-know-the-broad-outlines-of-literary-history/