Alongside the Last.FM model of personalised online radio (which I covered at some length and have cited in several other posts), Pandora provides an alternative based on different technology and classifications:
We take your input (artists, songs) and feedback (“I like this”, “I don’t like this”) and use the Music Genome Project™ to create stations that play songs that are musically similar to what you’ve told us. That’s it; only the music counts. We don’t care how popular the artist is, who’s backing them, and we don’t care which genre bin they usually belong in. Only the music matters. [Source]
Sounds good — and Pandora is now getting referenced in guides to online sources for finding new music, having more money behind it than Last.FM.
Pandora’s secret recipe is its grandly-titled Music Genome Project based on a set of elements which have been developed to map all the variations of western popular music. Real graduate-level labour goes into tagging each song on Pandora with these elemental characteristics, according to the description:
our team of thirty musician-analysts have been listening to music, one song at a time, studying and collecting literally hundreds of musical details on every song. It takes 20-30 minutes per song to capture all of the little details that give each recording its magical sound — melody, harmony, instrumentation, rhythm, vocals, lyrics… and more — close to 400 attributes! We continue this work every day to keep up with the incredible flow of great new music coming from studios, stadiums and garages around the country.
Nowhere can I find details of how many songs Pandora has analysed in its catalogue, but it’s perhaps unsurprising that their database appears not to have heard of artists a little outside the mainstream such as Bill Nelson or the Penguin Cafe Orchestra.
Pandora is in the process of expanding to cover to Latin music, though they confess they have yet to figure out how to tackle classical music.
I’ve spent an hour or two listening to Pandora. It was free to use when first launched, but now you have to register in order to listen to one of your stations for more than a limited period — and in order to register, you have to enter a Zip code to show that you’re in the US, which I’m not (OK, I could make up a Zip code, but…).
I was curious to find out the kinds of recommendations Pandora would make, whether I could find out more about the characteristics that make up the music genome, and how I could ‘train’ the playlisting program. Pandora provides just one window into its proprietary genome characteristics: for each song you hear, you can click to ask “Why did you play this song?” and Pandora will list the characteristics of the song that fit the profile it has of you.
I started by saying I like The Magnetic Fields and Neil Young, both of whom have covered a variety of styles and genres.
After playing The Magnetic Fields themselves, Pandora played The Concretes. Why?
Based on what we’ve told us so far, we’re playing this track because it features mild rhythmic syncopation, a vocal-centric aesthetic, major key tonality and many other similarities identified in the music genome project.
I decided to see if I could hear something with a less vocal-centric aesthetic, so I added The Ventures and Tangerine Dream. Next up I got Tangerine Dream, because it features
electronica roots, use of modal harmonies, a unique harmonic progression, acoustic hat and emphasis on instrumental performance,
followed by Kraftwerk (hmm, I wonder how many Tangerine Dream listeners have yet to discover Kraftwerk), featuring
electronica roots, acoustic hat, atmospheric production, a highly synthetic sonority and a dominant synth sound.
and then Harold Budd and Brian Eno —
electronic roots, a unique form, use of modal harmonies, acoustic hat and atmospheric production
followed by more ambient Eno with
electronic roots, a unique form, a unique harmonic progression, acoustic hat and atmospheric production.
To wake things up, I said I liked Child in Time by Deep Purple (I exaggerated slightly), which got me The Doors, on account of their
extensive vamping, minor key tonality, an electric guitar solo, prominent organ and many other similarities identified in the music genome project.
Hopefully you start to get the idea. Bearing in mind the supposedly clever analysis behind Pandora’s playlisting, it’s surprising how many of the connections it throws up are fairly conservative and ‘literal’ (the exceptions in my experience being tracks by Squarepusher and Cousin Silas, the latter being new to me). This Slate feature on Pandora makes the same observation, saying the matches provided for a Gwen Stefani track were “pretty much the same list you’d get from any music-savvy 12-year-old girl”.
Last.FM allows you as a listener to change how tightly or loosely the playlist is tied to your profile — in other words how adventurous and random you want to be — and perhaps Pandora will add something like that when it evolves beyond its current ‘version 2’ status.
I’m left with some specific questions about Pandora’s model, and some more general reservations. First, the questions.
- Given that the Music Genome characteristics are proprietary, and presumably will remain a jealously guarded piece of intellectual property, what scope is there for them to be validated, either by external musicologists or empirically (however you’d do that)? How do we know that they’re any better than another bunch of characteristics that someone could make up?
- I’d be interested to know whether the analysts’ recognition of particular artists influences the characteristics they assign? For example, do they recognise a Kraftwerk track and almost automatically tag it with “highly synthetic sonority” (is there a “moderately synthetic sonority” option?)? No matter how much training Pandora puts its analysts through, their previous familiarity with particular tracks and artists may well influence their classifications — has this been tested?
- Would Pandora allow users to add some of their own tags to the songs they hear? Then the structured Music Genome characteristics could co-exist with user-generated ‘folksonomy’ tags, in a similar way to that described by Colin Donald in his comment on my Web 2.0 post. Can Pandora’s system be opened up to user input in this way?
In his Slate feature, Martin Edlund makes an inspired comparison between Pandora and the prototype Global Jukebox system that renowned musicologist Alan Lomax worked on for the last decades of his life. Like the Music Genome Project, Lomax worked to unpick the ‘codes’ that combine to make different forms of music. Reflecting his expertise, the Global Jukebox is more concerned with folk and ‘world’ music than contemporary rock and pop. The way users interact it with it is also different, as Edlund writes:
Interacting with the Jukebox is like having a sophisticated conversation about music with someone much smarter and more cultured than yourself. The simplest way to use it is to pull up a map of the world, zoom in on particular regions and cultures, and listen to music clips. But Lomax wanted people to participate in the process of comparison and discovery as well. To that end, the Jukebox lets you compare and relate individual songs or entire musical cultures and trace traits across the globe.
The idea of a conversation seems more open than what Pandora offers — their recommendations are passed on like the words of an oracle, and are not open to question beyond the simple justifications quoted above. This not only means that Pandora has to keep the gargantuan task of maintaining its data to itself (rather than sharing it with others); it means that listeners may quickly lose trust in Pandora’s recommendations is they perceive they’ve received one or more duff ones.
Finally I have a more profound doubt about the genome model underlying Pandora in that it assumes the characteristics with which it tags tracks inhere in the music itself, independent of who is listening. In other words, that what one (trained) listener hears as a ‘unique form’ or ‘unique harmonic progression’ will have the same significance to other listeners, even though they have different experiences and training.
By coincidence I read yesterday something Brian Eno wrote in 01994 in response to a proposal to create a way of measuring musical compositions in terms of their complexity, and thus determining the degree to which people think it’s good or interesting music. He argued that complexity is very much in the eye (or ear, or mind) of the beholder — so an unaccompanied voice singing a simple song (say, June Tabor singing First Time Ever I Saw Your Face) can be complex because of the associations and ‘cultural undertones’ that some listeners bring to it.
[M]usic is actually a contingent combination of sounds whose emotional resonances are entirely dependent on the audience’s personal and shared histories as listeners. By “contingent” I mean that it could have been otherwise. Music didn’t have to consist of the elements and structures that it happens to consist of — and indeed it consists of quite other ones in other cultures, as anyone attending a concert of classical Thai music will soon realise… So complexity, I’m saying, has to be present, but present in the whole system — music and listener — as a system. If it’s just in the music (whatever complexity would mean in a purely
objective sense like that) it makes no difference to anyone.
Not all of this argument reads across directly to Pandora and the Music Genome Project: their use of hundreds of characteristics distinguishes their approach from one that relies on the single factor of complexity. But if you accept Eno’s first sentence above, you will not find Pandora’s potted rationales for its recommendations quite so convincing.