Technology has in many ways
democratized the music world, shifting the capacities for control and choice
away from the traditional holders. Now, more than ever, artists have the ability
to produce their art. New, less expensive recording technologies have allowed
any musician with a computer to produce an album from home. As a result, more
music is being recorded now than ever before. For musicians and listeners alike
this poses a new problem: how to sort through all the noise to connect the right
sound with the audience who is waiting to hear it?
Tim Westergren, founder of the Music Genome Project, offers a new alternative
for music recommendation. Similar to the mapping of the blueprint of life in the
Human Genome Project, The Music Genome Project attempts to unlock the secret to
music by deconstructing it into ~400 distinct characteristics; tonality,
harmonics, lyrics, instrumentation, etc. These musical characteristics or
"genes" are then used to make recommendations through a program called
Users can enter either songs or artists they like into Pandora which will create
a "station" of songs that share similar musical traits to their original entry.
In this way, Pandora matches audiences with the music based solely on sonic
qualities of one's own musical tastes. As claimed on their website, "In an
industry where less than 3% of all releases currently account for over 80% of
all revenue, [Pandora] is ideally positioned to unlock an enormous lost revenue
TMT recently talked to Tim Westergren to ask him about his process for
constructing this music recommendation system, his ambitions for it and the role
it serves in personal discovery.
Could you give us a little background information about
yourself; how you first got into music?
Well, I'm a musician and I first got into playing music when I was six. I was
trained as a jazz piano player. I went on to study composition and theory in
college and also spent quite a bit of time studying recording technology and
sound creation, some computer applications to music- not as a software
programmer but as a musician. After I graduated from college I spent about ten
years playing in rock bands, traveling around the country, experiencing the
challenges of being an independent musician. I also spent four years as a film
How did you make the jump from musician to founder of a music recommendation
I temporarily lost my mind is what happened! (laughs) Well, in 1999 I was a
film composer and writing music and running a recording studio. The idea [for
the Music Genome project] had come to me over the course of time, sort of
percolating in my head. In late 1999, it seemed like there was just so much
opportunity for online music to fundamentally change the way music is consumed
and potentially to change the way musicians are heard- to even invert the whole
pyramid that is the music business. I shared this idea, a classmate from
college, who had himself successfully launched and sold his own company. He had
a real can-do attitude about it and said, 'Hey, let's start a company.' A few
months later we raised some financing and just started in the beginning of 2000.
So, for me this was a big left hand turn - I was about to move to L.A. and get
completely buried in film composing.
Did you enter this business because you believed there was an untapped market
for unheard music?
That's exactly where I was coming from. I have long thought that the music
industry is a $100 billion business in $10 billion clothes and it's all because
nobody can find any music they like. That is the basic problem that everybody
faces. That's a problem for everyone: it's a problem for consumers; it's a
problem for musicians because they can't find their audience and it's a problem
for retailers because they have a hard time introducing their customers to new
music in their stores or online.
How do you think the Music Genome Project approaches this problem differently
than other music recommendation sites?
We have really taken a fundamentally different approach to recommendation.
There have been two modes of recommendations to date; one is called
collaborative filtering which is the kind Amazon uses. They compare your
purchase history to somebody else's and make predictions based on that. If you
buy a particular album they will search other purchases from everybody else who
has bought that album to try to decide the next most likely album for you to
buy. This system is pretty much a popularity contest. The other system is kind
of editorial, where companies have built large artist relational databases. They
will say, 'these two artists are like each other' and create these webs of
interconnection between artists. Outside of the traditional channels, which were
the experts; the DJs and record label president who recommend stuff, these have
been the two solutions available for the mass music market for a long time. We
have essentially foregone both of those and said, 'we are going to just really
understand the sound of the music and make recommendations that are purely based
on the sound of the music'. [We wanted] to do this in a really rigorous way
that's not the opinion of an editor, and not based on what other people like. It
is the first time any company has been focused on musical detail as a foundation
for musical discovery.
Could you elaborate on what parameters you use for
the genes that make up the music Genome Project. I understand that there are 400
defining characteristics you use to analyze each song.
They essentially cover all of the granular details of melody, harmony, rhythm,
form, compositional qualities and lyrics. I think of it as the primary colors,
the distinct elements [that make up a song]. For example, there are over 30
attributes that describe the voice alone; how much vibrato, range,
ornamentation, tone, performance. The sound of any voice, whether it's a Tuvan
throat singer or Mariah Carey, we have a basic collection of primary colors that
can describe it in one big continuum.
How long does it take to document each song?
About 20-30 minutes.
How many people do you have working on analyzing each song?
We have 32 analysts.
That seems like quite an ambitious task for so few people!
It's crazy! Keep in mind, we've been doing this for over six years now.
So the premise for your program is that musical taste should be based on the
musical components of a song?
Well, our premise is that that is a good way to start. There are of course
things we don't capture - there is a lot about music that has to do with social
and cultural phenomena, and personal components like when you first heard a
song, what's cool or not. There are a lot of elements we don't capture and we'd
never claim to be able to. I do think what we have is the most reliable for all
listeners to give you new music that you are going to like. I think that's been
borne out so far. We started our company by licensing our technology to other
companies. We have been picked time and time again by big retailers to be the
solution because our recommendations were consistently much better than other
alternatives like collaborative filtering and editorials.
You mentioned that you have licensed out the Genome to other companies, and I
have read that it has been featured in computer kiosks in Tower Records, Best
Buy and other music retailers. Obviously it is also the basis for your
website. What other mediums do you foresee the Genome expanding into?
Our long term ambition is to have this be the best listening experience,
anywhere, anytime. So that includes on a phone; any mobile device, when you're
traveling, when you're at home, you name it...we'd like you to access your
stations wherever you are.
Speaking of stations, Pandora is formatted to look like a personalized radio
station. What do you think are the main problems with traditional radio and how
will Pandora avoid these problems?
I guess there are a couple problems with radio. A radio station only gets to
send out one station that everybody has to listen to. Consumers are so different
it is very hard to program that one station to accommodate everyone, so it winds
up being the best radio for the largest number of people. I think this is part
of the reason for the homogenization of radio. If you are going for the largest
listener base you will tend to play the predictable, well worn hits and well
known artists. So radio has been very successful doing that for a long time, but
it is limited.
How will you avoid problems like payola?
In terms of payola, Pandora will never take a paid placement to decide
what's in a playlist. The recommendations we make are going to be based on the
genome, they will never be based on somebody buying the space. You heard it from
me here first. Never!
OK, so radio stations are influenced by promotion teams, advertisement budgets,
etc. So how does Pandora decide what goes into the music database?
It's kind of a big, ongoing R&D project. We have a couple of people who
spend all their time scanning every available music resource to pull new music.
It is of course the charts, we need to have what people know because that is
often what people use to start a station. We also look at small mail order
places, little indie labels, review sites, destination sites, CMJ (the College
Music Journal). Every potential place where we might some piece of music that
has found an audience, however small, we'll spend a lot of time trying to find
it. It is actually quite painstaking.
Do you see any flaws in this current system of music acquisition?
Well one of the issues is that because we do it with people we can't cover
everything. Our database is not as large as other databases.
You mentioned that people have been requesting music that wasn't in the
database. I wanted to ask you why there currently isn't classical or world
World music is coming- that's just a matter of priority. We had a huge
amount of music to get the North American and British catalogue done- what
American audiences would want. Now we're going after all other music. Classical
music is a different animal all together- it's a huge, huge beast. We're still
figuring out the best way to do it. Not only is it sonically different, it has
got a huge breadth but also the structure of the catalogue is different- 30
orchestras/symphonies doing the same piece. You're talking about a composer
driven and composition driven catalogue. It's a very different genome problem to
I read on your website that the genome for Latin music is also currently being
developed. Does each genre of music require a unique approach?
I would say each cluster of genres. The pop genome covers pop, rock, funk,
blues, folk, country; the mainstream genres.
How many overarching clusters are there?
There are four; pop, jazz, rap/hip-hop/electronica and world.
So when your team is analyzing these different clusters do they share a common
thread or do they each require unique approaches?
Every genome has a common genome- there is a common genome that crosses all
genres. But in some genres- World music for example, requires a much broader
palette of different instruments than are used is than pop music. Its doesn't
make sense to do all that work in pop music when it's redundant 99% of the time
so we adapt the template to closer match what the demands really are for that
genre of music. Rap, for example - there's more detail around lyrics in the rap
genome than there is in the pop genome because rap is so much more lyrically
focused. We'd focus on the literary and delivery- rhyme schemes, rhythm, and
wording; like how much profanity there is. It'd be like if you sat down and
said, 'I'm going to write 100 parameters that help describe what a rap sounds
like' you'd probably come up with something similar to what we did.
That's interesting; I didn't realize lyrical content was included in the genome.
We don't go too deep into it- but we go deeper in rap than we do in pop, for
Could you take the components, the most popular genes from the common genome in
all genres and construct the perfect pop song?
No, I don't think so. (laughs) What makes a piece of music good is hard to
quantify as a gene. It's not that detailed. I don't think there is a formula to
writing a hit song. If there was, everybody would be doing it. So, thank God for
Are there any other criticisms you have received since launching that you are
working on improving?
Over-repetition of songs has been one. Sometimes that's a problem with our
playlist system or because we don't have enough music in a particular part of
the universe. We are working hard on that. Music we don't have is our biggest
complaint right now. Another big comment we get is that we put artists on the
same playlist that should never be together. It's an interesting conversation
we've been having with people because we're sort-of anti that - our belief is
that if they sound similar we don't care if one looks cool and one doesn't or if
one has sold a lot of records and one didn't or one is something your dad likes
and one is something you like.
That idea seems to be one of the most fun and surprising parts of Pandora.
Exactly. I hope Pandora causes people to reevaluate that, reevaluate your
hatred of Neil Diamond - listen to his music.
Does your technology have the potential to reveal a truth about somebody that
they never realized before?
I hope that happens, I think it's true. People get so caught up in labels
that it doesn't even matter what the artist does and I think that everybody
would be better off if people started to dismiss those. I've heard a Neil Sedaka
tune I liked - there's never in a million years I would ever buy a Neil Sedaka
record just out of the blue, but maybe now...
Is it important for people to realize and deal with these biases on their own
rather than relying on a program. Does your technology give people a loophole
around self discovery?
Well, in a way I think you are discovering them on your own here. The thing
is, yes, this is a piece of software that you are using but it's not really
technology that is doing this - its musicians really. Pandora is just a way to
take the knowledge that these musicians have and make it available to lots of
people at the same time. So, sure, it is technically a piece of software but the
recommendations are really human recommendations when you think about it. The
whole genesis came about because I became pretty good at figuring out what
people like in music; as a composer, in writing music that they would like. I
became pretty good at recommending stuff and that's what Pandora is supposed to
be; the small, indie, record store owner who knows you personally, knows what
you like and has this encyclopedic knowledge of music. So, I wouldn't really
consider what we have as a technology solution per se. It uses technology but it
uses musicians' ears more than anything. It's a musician who makes the
recommendation to you- they just do it via a piece of software.
Do you think Pandora will broaden tastes or deepen already established tastes?
Currently, it's primarily to deepen but we have the ability to allow you to
specify things as you create your playlist about the music you like. It can
build new playlists, not just based on the overall song but a particular element
of it; the sound of the voice, the guitar playing, the lyrical content or the
harmonic structure if you're more musically inclined. I think that's going to be
a broadener. If you listen to a Rolling Stones tune and you hear a sitar part
that you like- you click that and all of a sudden you are into Indian music or
Western music that uses sitar more and more. You have this whole world open up
to you- so that's how we hope we will be a real broadener.
You launched the website a few months ago. What has the response been like
We came out on the web just a couple months ago and the response has been
unbelievably, overwhelmingly positive...we've got 100,000s of users. Recently
we passed the millionth station created from around 70,000 artist or song feeds.