Now that I have an office job, I’m spending a lot of time under headphones while I correct people’s grammar. It’s a good opportunity to explore the outer reaches of my music tastes. The office has some networked iTunes libraries heavy on the Pitchfork 500, and I have whatever I’m bringing from home. I’ve also been making my first serious adventure with internet radio. I arbitrarily picked Pandora because they have a free iPhone app. The web version is nothing to write home about design-wise, but the iPhone version is fun, and over wi-fi there are none of the buffering delays that have kept me from enjoying internet radio in the past.
Computers can pseudorandomly generate playlists. That’s also what human DJs do. So does that mean that human DJs could someday be replaced entirely by laptops? So far, nothing an individual computer can do comes close. Computers can take a batch of songs and play them in a random order, and that can sound okay when you carefully pick the batch of songs. But you, the human making the playlist, are still doing most of the intellectual heavy lifting.
Internet radio stations come closer to mimicking human DJs than standalone computer software like iTunes. Internet radio stations like Pandora try to sequence tracks semi-randomly, semi-automatically in a way that you’ll enjoy. Internet radio draws on big databases filled with metadata on what makes one song like or unlike another song.
A good database like Pandora’s isn’t a simple spreadsheet layout. Pandora draws on a giant online database called the Music Genome Project. It sprawls along dozens of information dimensions and was assembled laboriously by hand by a panel of musicians and DJs and programmers and executives.Instead of grouping tracks into unambiguous genres like iTunes annoyingly does, Pandora uses non-exclusive tags, in a process that I imagine works a lot like Delicious or Flickr tagging.
To listen to Pandora, you start by typing in the name of a song, artist or genre. Pandora then has some quasi-random process by which it serves up tracks that match as many tags with your starting point as possible. You can give each song a thumbs up or thumbs down, and as it accumulates your votes it makes better guesses as to what you’ll like. Sometimes these guesses are on-point, and sometimes they miss completely.
I started my Pandora adventure by giving it something easy: “Jesus Walks” by Kanye West. It played me some of Kanye’s hits, then some of his more obscure tracks, then tracks by other artists where he does guest verses. Then it segued into Common and Jay-Z. So far, so good. I tried Gang Starr, and Pandora was similarly successful, lining up a string of golden age hip-hop classics. Then I tried giving it Thelonious Monk solo piano. It did better than I expected, playing quality fifties bebop, heavy on Sonny Rollins. Monk’s music might be peculiar, but I guess there’s so much jazz writing out there that the Pandora people had plenty of help giving nuance to their descriptive tags.
So Pandora does great with well-defined genres and styles. But it doesn’t do so hot with outliers and edge cases. In response to Herbie Hancock’s “Rockit”, it serves up “Cantaloupe Island,” okay, but then it follows with some lame New Jack Swing, then, jarringly, “Rock Me Amadeus” by Falco. I wanted early hip-hop and electro. After its few weak stabs at jazz, Pandora gave me synth pop. I understand why it did that, but it’s not what I was looking for.
Pandora did even worse with Björk.
In response to “Jóga,” Pandora served up one so-so Goldfrapp song and then a bunch of bland electronica with anonymous-sounding female vocals. Some of Björk’s music resembles the stuff you hear in the lobbies of hip hotels. Some of it wildly doesn’t. If you’re after artists who sound like Björk, there aren’t any. There are artists who sound totally unlike her but appeal to the same people, namely me, because they’re weird and experimental sometimes and make you want to dance sometimes.
So Pandora fails at Björk. In fairness to the people behind Pandora, her supposed genre or style is a terrible predictor of whether you’ll like her or not. Plenty of people I know who love dark, moody electronica can’t stand Björk. Most of the jazz musicians I know adore her. Every jazz group I’ve ever been in has played at least one of her tunes. NYC has a seventeen-piece big band that plays nothing but her music. Björk might seem like an odd fit for hip-hop lovers, but Timbaland sampled “Jóga” on a Missy Elliot remix and the Wu-Tang Clan namechecks her.
My initial reaction to Björk as a college student was: immediate, strong dislike. It took me years of constant exposure from my sister and friends to change my mind. Now Björk occupies a prominent spot close to the center of my musical affections, close to Coltrane. Who, now that I think about it, I also wasn’t too wild about on the first hearing. People are weird and unpredictable in the evolution of our likes and dislikes.
The computer world has shown great ingenuity in the past few decades at creating bigger and more intricate databases. But there are intrinsic limits to the kinds of information a database can represent. Computers are good at two things: strictly following unambiguous instructions, and producing total randomness (or pseudorandomness that’s close enough to random to fool us humans.) Computers aren’t so good at blending the random and the structured.
I think we can expect computer DJs to continue to get better at predicting our predictable tastes, and to not make much progress on the unpredictable ones. Pandora might do a better job making playlists than people who aren’t very serious about music, but I don’t think that computers will ever have anything on professional club DJs. There are too many variables to quantify.
What’s most interesting to me about quasi-random database managers like Itunes shuffle and Pandora is the way they inform my own selectional process. I love shuffling within my own carefully cultivated playlists. The real value is the way the computer gives me unexpected items to choose from, which I then apply human emotional intelligence to.
My friends in the academic music world are very interested in algorithmic composition right now, where the computer generates semi-random strings of notes within certain rule sets. I find nearly all of these compositions to be unlistenable. I do enjoy trying out random MIDI sequences as a source of inspiration, but the real music-making happens in my rejection of most of those sequences, and my editing and adaptation of the best ones. So it goes at the playlist level too.