About half a dozen people circle a large conference room in Google’s New York office, reviewing a series of color-coded spreadsheets projected on a wall behind them. Peter Asbill, who oversees editorial for Google Play Music after the company acquired his music start-up, Songza, last summer, tells them, "Let's jump into some stations."
There are currently 86 different “work-out” playlists, each one made and revised by human beings.
The process is reminiscent of the days of when music fans might have carefully crafted a mixed cassette tape or CD for a friend or crush. In today’s digital music world, an increasing number of music streaming sites want listeners to outsource those playlist-making responsibilities to them, competing for the ever-increasing time consumers spend with earbuds in their ears.
But perhaps surprisingly in this digital age, even when playlists go corporate, human beings are still key to their success.
For Asbill, the key difference between yesterday and today’s process is data. The group focuses on their “Performance-Enhancing Pop” playlist. Data collected from users shows that the playlist gets selected frequently, but its “satisfaction score, a proprietary score that’s a mix of skips and thumbs up, thumbs down, and listening time,” is just okay.
The worst performing song on the playlist is "Honey," by Mariah Carey, which the group decides is likely too slow compared with others on the list. A Kenny Loggins song also isn’t preforming well. While its tempo is faster, the team thinks it’s stylistically out of place with other, more popular songs, like Lady Gaga’s "Just Dance."
In the end, the consenus is that Kenny and Mariah should probably be cut from this mix, along with Britney Spears, “I’m a Slave 4 U,” which the group thinks should moved to their “Sexy Sweaty Workout” list.
But why is this human touch necessary in 2015? Why not take this data and let machines and algorithms do this work? Asbill says Google Play Music does use algorithms when people want to pick an artist or song and have site suggest similar music. But he says playlists made by humans can be creative, funny, or unique to a specific situation, activity or feeling.
“It always requires further investigation,” he says. “If it’s a new song, is it too early so that not enough people know about it, and like it yet, and will it do better over time? Is it a song that’s a great song, it’s just on the wrong playlist, or have we given the playlist the wrong title so that we’re not setting up people’s expectations for what they hear on the other end?”
Pandora, an early entry into music streaming, also relies on people for its music curation, but in a very different way from Google. Both sites use algorithms to create a playlists based on a song or artist the listener chooses. However unlike Google, which has machines analyze those songs, Pandora uses people.
“It really takes a human,” says Steve Hogan, Pandora’s music operations manager. “For example, a human can listen to a song and pick out there’s a trumpet, there’s an acoustic bass. This is something the human ear can do in a fraction of a second.”
A team of roughly 30 people, most of them active musicians, analyze 10,000 songs a month, tagging 200 to 400 traits for each song, answering questions like, "How much does the electric guitar dominate this song?"
“That would be on a five-point scale,” Hogan explains. “Beyond that, we want to know, 'how much distortion effect is on that guitar on a five-point scale? What is that guitar doing, how much of it is strummed rhythm guitar?'”
The algorithms then draw on those descriptions and tags to build playlists of similar songs for listeners. In other words, machines offer scale for Pandora, but it’s humans who bring the quality.