The Scoop on Pandora for the iPhone and Other Platforms: Tim Westergren Speaks at Boston’s Apple Store

11/18/08Follow @wroush

(Page 2 of 3)

weird, I really like this music.’ Then I realized ‘Oh, it’s my Pandora station.’ Then a song came on that I didn’t like, and I just skipped it. It’s amazing how emotionally different it is. In a car, you’re used to being captive to broadcast stations, and all of a sudden you have this personal listening experience.”

The feeling that Pandora knows you and your musical tastes is only possible thanks to the years of work the company has put into its catalog of music ratings, called the Music Genome Project. Since Pandora’s launch in 2000, the company’s staff of 50 trained musicologists has classified 600,000 songs, and continues to add about 10,000 songs per month. It’s an incredibly labor-intensive process, Westergren says. “For a simple, three-minute pop song, it might take 15 minutes” for a Pandora musicologist to rate a song along all 400 parameters. “A symphony might take two hours.”

Having recently visited and profiled The Echo Nest, a Somerville software company with music-recommendation software that automatically—and very quickly—assesses a song’s acoustical qualities, I asked Westergren how Pandora can keep up with all of the music coming out these days, given its slow, meticulous, human-dependent classification method.

“It’s an absurd approach,” Westergren agreed jokingly. “But it’s the only way to do it, we think. It also has one really distinct and unique advantage, which is that it’s blind to popularity. If a garage band’s music goes into Pandora, it gets an equal shot to Bruce Springsteen. Every other system is really about collaborative filtering–that people who like this also like this. You don’t get recognized on Amazon if you’re not already somewhat popular.”

Actually, The Echo Nest’s software is an exception to the collaborative filtering phenomenon—but I didn’t get to pursue that with Westergren. He followed up, however, with another important point: while automated rating software may churn through more music, not all of that music is worth a human’s listener’s time. “Our goal is not to be all-inclusive,” Westergren says. “We want to try and find just the best music. I can say this as a musician: most music that’s published is not very good. Maybe 5 or 10 percent is actually ready for prime time. That’s the stuff we want to find, and I think we have an operation big enough to do this.”

Pandora makes money on the whole operation by selling ads alongside its Web-based console; the ads are customized according to users’ gender, age, zip code, and musical taste. The company also provides links from its songs to Amazon and iTunes, and earns a small commission on all song purchases. But while they’re small, the commissions can add up—especially given what Westergren calls the “crazy” conversion rate on the site. “Forty percent of listeners end up … Next Page »

Wade Roush is a contributing editor at Xconomy. Follow @wroush

Single Page Currently on Page: 1 2 3 previous page

By posting a comment, you agree to our terms and conditions.

  • http://www.HearWhere.com Pete

    Great article Wade,

    One note, though the Echo Nest is currently the darling of the ‘acoustic matching’ world, they are not the only ones who are looking more at the acoustic attributes rather than collaborative filtering.

    MusicIP has been doing this for years, and their music recognition algorithm (fingerprint) is used by MusicBrainz.

    Check them out http://www.MusicIP.com, and if you’re a fan of live music, be sure to check out my site http://www.HearWhere.com