with CLIPREVIEWED learn the articleLove Your Spotify, Apple Music Recommendations? Thank a Pandora Music Analyst
Before Pandora’s personalized internet radio service launched in 2005, there was Savage Beast Technologies. From 2000 to 2004, the company focused on building out its music-recommendation technology, the Music Genome Project. Pioneered by founder Tim Westergren, the Music Genome Project had humans, in the form of skilled musicians, listening to songs to uncover and annotate their musical attributes.
Remarkably, music analysts are still listening to songs today. Many aspects of how streaming music platforms provide song recommendations has shifted to computers and machine learning, but not all. There are still analysts doing the same work they have for the last 20 years. We talked to several analysts, past and present, to get a sense of how they work today, 20 years after the Music Genome Project’s inception.
Human or Machine Recommendations?
“I think we fit in the way we always have, which is as a total anomaly,” says Pandora Music Analyst Scott Rosenberg. “As far as I know, we’re still the only service where the human component is still so central. At least at the scale we’re working at.
“Whenever we hear from people in the Music Science department—the folks working on machine learning—they are all so flipped out that they have this vast, multi-million song repository of human-generated analysis to draw on. There is no other set of musical analysis data that even remotely approaches what we’ve created.”
That vast repository is possible because of the amount of time Pandora has been doing the work. Steve Hogan has seen the experiment of human-led analysis grow since he started with the company in September 2000. He is now Pandora’s Director of Music Analysis.
Twenty years on, Hogan still believes music analysts are Pandora’s backbone, “but it was especially true in the early Savage Beast days, when our product was so squarely focused on the Music Genome Project,” he says. “The Music Genome Project remains the foundation, but our product has diversified, as has our set of recommendation strategies.”
There has been debate over whether computers could ever be as prolific at song recommendations as humans. In truth, humans still play a large role in music discovery everywhere. Discover Weekly, a weekly automated playlist from Spotify, is regularly heralded as a brillant discovery source, but it isn’t purely machine-based. Songs are crowdsourced from listeners and their libraries and then smartly and quickly analyzed with the help of computers.
Hogan suggests Pandora’s initial work in this area could also be credited with spurring a focus on personalization at Spotify, Apple Music, and others streaming services.
“It forced them to offer individualized, algorithmically driven playlist features,” he says. “Although, for those two services, I think the on-demand aspect is more the focus. Maybe I’m biased, but I believe we are still the best in the business in the area of recommendation and personalized listening, but there’s still plenty of innovation left to do there.”
Becoming a Music Analyst
Music analyst is a somewhat obscure profession. It involves listening to songs and categorizing them based on chord progressions, tempo, genre influences, vocal qualities, and more, but what else goes into it?
“The most challenging part of analyzing was developing the confidence that I understood the genome the same way as the other 30 or so music analysts. This thing can only work if we are very consistent across the entire team,” says Hogan. “I very quickly encountered every shade of gray in the music I was hearing, which made it difficult to fit neatly into the genres we had available to us. At the beginning it was very slow going, and there were seemingly infinite questions to be resolved from all the music analysts.”
It has changed over the years, but in the early days, music analysts reported for duty at an office in Oakland, California. Each week, on music release day, the most popular albums were purchased for the group and stacked on tables. Music analysts made a selection, donned a pair of headphones, popped in a CD, and sat at a desktop computer to listen and annotate.
“I remember analyzing Britney Spears’ song ‘Toxic,'” says Nate Query, bassist for the band The Decemberists. “It was a weird one to try to analyze because there’s so many instruments and all this stuff going on, but I was like, wow that’s a jam. I still like that song a lot to this day.”
Query met Steve Hogan playing a wedding gig one weekend, and Hogan invited him to apply for a job as an analyst.
“It sounded way better than answering the phones at a law firm so I said sure,” says Query. “It turned out Tim Westergren was the head guy there at that time, and I knew him from the 90s when I played in a hippie band that did tons of shows with his band.
“When I auditioned for the job they gave me a couple songs to listen to, one of which was ‘Pump It Up’ by Elvis Costello, and said you need to answer these questions about it.”
Query was only an analyst for a few months before The Decemberists began touring for their album Picaresque full time. Nate Brenner of the group Tune-Yards also worked as an analyst and within a few years, he was able to take the job on the road while touring, analyzing songs before shows and at random times in his schedule.
“At first it was challenging learning how to analyze correctly based on the manual and the system that was in place,” says Brenner. “Later it was hard to not get into an auto-pilot mode and to remember to slow down and pay attention to the details of each song. It was also hard to listen to music all the time and not let it affect my own creativity.”
The Legacy of Music Analysts
The Music Genome Project initially grew out of a desire to figure out the musical connection between songs. It was a method of discovery that ignored popularity bias or the stigma of genres, and opened up a way for consumers to find music beyond what was being hyped by a record label or radio DJ.
“At the time it felt really exciting that Savage Beast was willing to kind of treat all music equally and not be as much of a gatekeeper based on some guy that’s been in the music industry for a long time. And it is a cool way to find music,” says Nate Query.
Pandora’s music analysts were key to that mission and gave life to fresh music discovery. Even though the music service’s influence has waned over the last few years (Sirius XM acquired it in 2018 for $3.5 billion) and machine learning has crept in everywhere, the volume and dedication to analyzing songs makes the Music Genome Project an indispensable resource.
Having a consistent team “is part of what makes the data all the more valuable,” says Scott Rosenberg. “If you’re changing analysts all the time, the standards shift and it’s harder to maintain a baseline.”
People like Steve Hogan have been around since the Nolanic Scrolls—the nickname the team gave to its analysis manual, written by Nolan Gasser, the company’s first musicologist. Having that type of institutional knowledge “means that we’re able to keep the whole project on track and maintain the original inspiration about mapping a musical genome,” Rosenberg says.
keyword: Love Your Spotify, Apple Music Recommendations? Thank a Pandora Music AnalystLove Your Spotify, Apple Music Recommendations? Thank a Pandora Music AnalystLove Your Spotify, Apple Music Recommendations? Thank a Pandora Music Analyst