In this extended version of the talk from our new issue, we speak with Tom Conrad, the CTO and Executive VP of Product at Pandora. Why does it take so much input from so many sources for the company to build perfect playlists?
Photo by Toby Burditt
"Hundreds of musical attributes are used to describe songs. Twenty-six alone describe vocal performance: how breathy or gravelly it is, how much falsetto is used, and more. That's too much nuance for software on its own. So we hired humans--specifically, professional musicians with four-year degrees in music. Every day, they come into the office, put on headphones, and listen to every song that will ultimately play on Pandora, listing as many as 450 different attributes a pop. When combined with users' past listening habits, those detailed fingerprints enable our algorithms to discern, for example, that when you say you like the Beatles, you're talking about their earlier sound and not their later one."
Fast Company: How does Pandora find all its music?
Tom Conrad: We have a team of curators of the Music Genome Project whose goal is to scour the Earth for every song that's ready to be played on radio. A key driver is the rare occasion when our users try to create a station and we don't have the artist or song they're looking for--that's a signal to us that there's demand for that artist. They use every kind of public forum of music criticism as a tool for understanding what's out, ranging from underground blogs about emerging jazz artists to Pitchfork on the indie music side. They also watch all of the charts, from pop to contemporary Christian music. We also have an independent submission process where artists who may be literally just recording in their garage can come to Pandora and upload tracks for consideration. We get hundreds of submissions like that a week and about half the music submitted makes it onto the service.
Why is a song's "musical DNA" the best Pandora?
The one thing you can know absolutely about a piece of music is what musicological constructs are at work in a song. If you can understand that, you can, by extension, find other songs that are musicologically similar. Which means songs don't have to be popular or well-known to be able to participate in the Music Genome Project system, since it's based on underlying musical attributes. And oh, by the way, it turns out peoples' musical preferences are hugely influenced by what the music sounds like.
What is so special about a song's "musical DNA"?
When we catalog songs, we don't just ask, "Are there guitars?" It's also about what kind of guitar, how the guitar's played, or what role the guitar plays. And that allows us to begin to build a service like Pandora. The way I like to describe it is the Music Genome Project provides an initial hypothesis about music that might work for you in the context that you've asked for. So if you say the Beatles, the Music Genome Project is going to say here are some songs that are likely to be successful with this listener because of their musical similarities to the Beatles' catalog.
But not all songs are always going to be successful.
That's where a third algorithm joins the mix of the Music Genome Project, which is a completely level playing field, and your own thumb feedback that's pushing a station in one direction or another. We also cluster the billions of pieces of thumb data that we get back from our users to make an even more informed projection about what songs might work for you, which allows us to bring a kind of wisdom of the crowd into the experience. For example, it might be that for a band like Massive Attack, the Music Genome Project might put forth a song from Jennifer Lopez, because there are some reasonable musicological markers that connect those two artists. But if the listening audience for Massive Attack comes along and says, "We can't quite handle the fact that you're proposing Jennifer Lopez to us," we can use that to dial down Jennifer Lopez for that audience.