What Streaming Music Algorithms Really Measure
photo by Kaboompics / Karolina from Pexels
It’s no secret that streaming music services are collecting data about us and using it to serve up other music we might like. I’m a diehard Spotify user, and they offer this feature in a few ways—there’s the Release Radar playlist, which curates new releases from artists you listen to often or might like, and Discover Weekly, which pulls in artists you may or may not know but are similar to ones in your universe. Then there’s the “radio station” option, originally pioneered by Pandora.
I left Pandora a long, long time ago because I found its suggestions vapid, poorly curated, and lazy. But there’s an even bigger debate happening among music lovers about the validity and quality of algorithms in any service and their ability to truly pinpoint our musical tastes.
A recent (informal) survey of several friends who are avid users of streaming services all pointed to a similar sentiment: Algorithms are crap. As one friend described, “It wants to pigeonhole me as either a get-off-my-lawn Freedom Rocker or a 19-year-old young woman.” The problem with algorithms is the same problem with generalizations—even 19-year-old women don’t all listen to the same kind of thing. And yet nuance is hard for a machine to learn.
My experience is similar to theirs—out of every curated playlist I pour over, there are maybe two or three songs that resonate. But I keep going back to those playlists because it’s kind of like eating trail mix that has dark chocolate bits inside. Sure, most handfuls are going to deliver raisins and sunflower seeds and little chunks of dried fruit—but once in a while, you’re going to find a few pieces of delicious, creamy chocolate. So you keep plunging your hand in the bag.
Streaming-music platforms may give us access to a plethora of choices and options in mere seconds, but why is it so hard to pinpoint our musical tastes? I went back and listened to my Discover Weekly playlist and tried to analyze each piece. They were all mostly about being strong and overcoming hard things—probably because I’ve been listening to a lot of what some might call motivational material lately, trying to psych myself into being strong enough to deal with a big life event that’s in the works.
And then it hit me. Algorithms aren’t measuring our musical likes and dislikes so much as they are mapping our emotional states at any given point. They’re trying to capture the mood, melody, tone, and overall feel of each piece we listen to and then spit similar songs back at us, mirroring what they think we’re feeling.
The problem with this is that we’re humans. Our moods change, all the damn time. Most of us have very diversified music tastes, and we listen based on how we feel. On a foggy early-spring New England day, I have a strong penchant for Andrew Bird. But don’t play me a song off Noble Beast on a hot July day for the love of god.
With all the buzz about robots and their impending takeover of all the jobs, we can rest assured that predicting human moods and therefore musical tastes is probably best left to us humans. Machine-generated melodies just don’t quite get it right.
Ashley Daigneault knew she was a writer before she left kindergarten and has a particular
love for writing about tech, literature, music, and politics. She is currently the VP at Caster
Communications, a full-service tech PR and social-media firm, and works with B2B and
B2C tech brands. She lives in New England with her family, which includes kids and dogs
who think they are kids.