Machines to Listen for You

A recent story on NPR’s Morning Edition describes a new software system to analyze songs for hit potential. Quoth Music Intelligence Solutions CEO David Meredith:

“[It’s] a series of algorithms that we use to look at what’s the potential of a song to be sticky with a listener,” Meredith says. “To have those patterns in the music that would correspond with what human brain waves would find pleasing.”

Meredith’s software uses algorithms to model how a brain finds “pleasure.” Even putting aside the strange attribution of taste to brain waves (as opposed to just the brain, or—I don’t know—the person?), this model has a lot in it to talk about. First off, the idea that pleasure is caused by “patterns in the music.”

What kind of patterns is he talking about?

Meredith says his software found that hits have certain common patterns of rhythm, harmony, chord progression, length and lyrics.

The choosing of parameters, even seemingly innocuous ones like “rhythm” and “harmony,” is an interpretive move. These parameters define the way the software can “hear,” grouping or separating various songs in contingent ways. Of course the story is vague about how exactly the software parses patterns in “lyrics” or “length”—you wouldn’t want to give up any trade secrets. But let’s take length as an easy one.

How would you incorporate length into an evaluation of a song’s “hit score?” It’s presumably a single number. Is shorter better? Maybe the ideal length is determined by a relationship to the various other scores? How do you extract a hit rating from it? By embedding these questions in software, Music Intelligence Solutions obscures them. The software must have a methodology, just as the music critic would, but since it outputs a score—7.6 out of ten is apparently “good for a platinum rating”—this methodology goes unexamined.

David Bell, of the hip-hop duo the Block Scholars, paid $90 to use it.
“To me, it’s an unbiased validation of your music,” Bell says. “It’s not your family turning around and saying, ‘Oh, you got a great song.’”

The computer told Bell he had a 7.1 — good, but not great. So he went back to the studio and remixed. He got his score up to 7.6 — good for a platinum rating. He could hold his head up.

The software configures musical production in a particular and contingent way. Bell produces a song, the machine evaluates it, he remixes, and the machine reevaluates. In the context of evaluative software, the labor of remixing is a negotiation with the machine.

The assumption is that the machine that turns music into numbers, processes them, and gives back a number, must be “unbiased.” By displacing evaluation from people to “a series of algorithms,” Music Intelligence Solutions banks on the obscuring power of technology: “your family” has bias, but a machine does not. Of course, the workings of the machine are entirely sculpted by bias. Each decision about what matters in a song is tweaked and informed by the programmers. The algorithms do not make themselves, nor do they decide which musical traits are significant.

Objectivity is produced through the use of numbers and software interfaces. Technologically mediated bias becomes objective evaluation. Software like Music Intelligence’s UPlaya propagates a certain view of what “hits” are, and even exceptions to the rule are integrated into a particular musical viewpoint:

It doesn’t surprise New Yorker music critic Sasha Frere-Jones that a computer can predict hits, but he says it can’t predict all the hits. Sometimes, songs come along that don’t fit the mold.

Songs that are popular in spite of their evaluation results (Frere-Jones uses the example of “Da Da Da” by Trio) become songs “that don’t fit the mold.” Now, not fitting the mold is another story about how songs become popular. On one side, we have software to account for certain kinds of success (like “I Gotta Feeling” by The Black Eyed Peas—8.9 out of 10), and on the other, a popular cultural myth about the appeal of mold-breaking. Evaluative software recasts songs that don’t meet its model as part of another cultural narrative.

Is this something to worry about? Probably not. For the integration of taste into business decisions, numbers can be very useful—they turn the the qualitative into the quantitative, with its sheen of objectivity. For listeners, though, I find it hard to imagine that you’ll be seeing ads for “The new Britney Spears: 9.1 out of 10 on the hit scale! You’ll Love It!” Actually, I take that back. You might really see these things, like impact factors on science journals, or prominent magazine reviews on CDs. If they actually use these in promotion, and not just in house, that would be both totally wild and fodder for like a million more blog posts.