For those who haven’t had the pleasure yet, Pandora is an online personal radio station site that creates playlists for listeners based on seed songs, artists, or albums. Enter The Beatles, and you’ll get songs by them as well as solo efforts by John and Paul, tracks from the Rolling Stones, Queen, and so on. Enter in Kanye West, and you’ll hear his music as well as that of related artists like Drake, Jay-Z, Rihanna, etc… Kinda magical.
Pandora's Playlist Panel
At a recent talk, company founder Tim Westergren took attendees behind the curtain and explained how the system works. Like most people, I expected that underlying all of this brilliance would be some sort of semi-scentient HAL/The Matrix/SkyNet style server farm, churning away at terabytes of mp3s and determining through soulless computer analysis that John Coltrane should be followed by Miles Davis.
Instead, what I learned is that, at the heart of Pandora’s system are a staff of forty or so trained musicians that come to work everyday to listen to music. Really! Individual tracks are listened to and scored against four hundred attributes (they call them “genes”) like tempo, instrumentation, and key changes. Once a track has been “gene sequenced” its unique collection of attributes can then be matched to other tracks with similar genes, and the seemingly “magical” playlists can be created.
The recommendations and pairings created by this method work pretty well on their own and are further augmented by feedback of the site’s listeners who, by clicking thumbs up or down buttons during songs, can help to refine the system’s understanding of which tracks should and shouldn’t be paired with which.
Apparently the engine is working exceedingly well. According to Westergren, 90% of Pandora’s catalog of 750,000 tracks are played each month by its millions of listeners. Given that in the 50 year history of Billboard’s charts, fewer than 100,000 songs have ever appeared, the exposure of that many tracks to such a large audience is an object lesson in the internet’s famous “long tail.” Moreover, because the engine is blind to pre-concieved ideas of what constitutes a match, unexpected and sometimes amusing relationships sometimes present themselves to the delight (and occasional dread – AUDIO LINK) of consumers.
NOTE: This is the last in what has turned out to be a trio of posts about Pandora (I promise!) that came about as a result of a lecture Pandora founder Tim Westergren gave on the service and its underlying engine, the “Music Genome Project” (MGP) on March 21st, 2010, here in New York. Click here and here to see the related articles.
