An individual’s relationship to music an album, a melody, an artist is often beyond the scope of verbal description. But can it be quantified in other ways? This was one of the questions touched on by Pandora CTO and Berkeley alum, Chris Martin, in an Engineers Week panel on the highly successful internet radio company. In an increasingly competitive streaming market, Pandora has maintained its hold by creating a unique listening experience. Martin’s talk forces us to wonder: what can’t be quantified in our data-driven age?
Martin began his talk by highlighting Pandora’s focus on consistency. It is one of few sites and services that is always online, setting aside no downtime for service improvements. Much of the backend code used today has remained untouched from when it was first implemented ten years ago. This is a feature of Pandora’s emphasis on quality work. “I would rather have forty hours of really good code in a week, rather than fifty or sixty hours of sub-par material,” shares Martin. He is quick to compare Pandora’s policies with their heavyweight competitors like Google and Apple, who woo their engineers with perks but may receive weaker output from them. The longer hours are inefficient and unreasonable, says Martin. At Pandora, “we won’t serve you dinner at work — we want you to eat dinner at home, with your family.”
But many of Pandora’s decisions lie on a more existential plane. In a conversation on data applications, Martin poses a deceptively simple question: “What is a fan?” The answer is not clear. How can we quantify when an individual feels the emotional, psychological, aesthetic ties to an artist that force them to self-identify as a fan? Is it a user who creates a station in the artist’s name? Or simply someone who “thumbs up”s their songs while listening? These are the questions Pandora must wrestle with as it seeks to advance the listening experience.
The quantification of taste follows a long trajectory of scientific rationalism largely inherited from the European Enlightenment. The belief is fundamentally in the scientific method, in empiricism — the idea that the numbers don’t lie; only what is shown to be true through repeated tests and analysis can be confirmed as fact. These quantified facts constitute the basis of knowledge.
In many ways, Pandora’s technology buys into this conceit. As Martin shared, Pandora owns the largest data set on listeners’ tastes from its “thumbs-up” and “-down” buttons. This user feedback, Martin argues, is what artists need to understand their fans holistically. Furthermore, the findings in the data can be used to enhance the relationship between the artist and the listener. This is the reasoning behind Amp, Pandora’s new service through which artists can deliver targeted videos to fans in the middle of carefully selected playlists. These shout-outs are meant as a form of advertising but also of intimacy between giver and receiver. How is this different from terrestrial radio? As Martin saw it, when Kanye West gives a shout-out on a radio channel, the channel is betting on finding an audience composed of some Kanye West fans. When artists use Amp, they will know that their listener is a fan, lowering the chances of misfiring and increasing the possibility of a connection.
The improved precision can be particularly useful for less-recognized artists, who may not be allocated similar platforms through traditional radio. Pandora has seen itself from the beginning as a high-profile streaming opportunity for emerging artists. Of the 125,000 artists in Pandora’s library, all of them are consistently in play, which Martin contrasts to radio stations (Berkeley’s own KALX excluded) that tend to play the same ten artists on repeat throughout the day. Martin also cites that many artists tend to have more Pandora stations in their name than followers on Twitter.
Music Genome Project
While Amp functions on the belief that data can be used to quantify something as complex as feeling or taste, Pandora’s basic operations rely on a more present human element. As stated on the website, Pandora’s music recommendation algorithm, known as the Music Genome Project, does not use machine learning or predictive modeling to assign music to users. Instead of feeding data into an algorithm that would classify and cluster the music, spitting out distinct values based on quantitative analysis, the project employs a much more qualitative approach. Over 450 musical traits are assigned by humans, not computers, to each song. Pandora’s employed musicians spend approximately half an hour analyzing each track. While quantitative data such as “Number of Thumbs Up” or “Duration of Listening” are more easily processed by the algorithm, Pandora believes that musical traits such as “Smooth or Silky” or “Level of Groove” are best assigned by humans.
In our post-Enlightenment society, Pandora’s vision for the listening experience continues to lie somewhere between the human and the mechanical. As Chris Martin puts it, the ideal listening experience is one where “the music plays for you whenever you want to hear it — this sort of serendipitous experience.” We may continue to be modeled and quantified, but as Martin’s talk at Berkeley revealed, Pandora’s use of technology and empiricism seeks not to reduce the human experience, but heighten it.