http://www.theglobeandmail.com/servlet/story/RTGAM.20051110.wxhits10/BNStory/Business/
Tragically, it's not hip
By GRANT ROBERTSON
Thursday, November 10, 2005 Posted at 5:30 AM EST
>From Thursday's Globe and Mail
They know what songs you like, even before you like them.
Record labels spend millions of dollars each year trying to predict
what singles will top the charts and which ingredients make a hit
single.
Now, two Massachusetts Institute of Technology PhD grads believe they
have cracked the code.
After years of crunching data, Brian Whitman and Tristan Jehan have
devised a computer program that listens to a song, then predicts how
humans will react to it.
The response is so specific at times that it can forecast how a single
will perform on the charts and spit out a review, guessing what words
will be used to describe it, from "sexy to romantic to loud and
upbeat," Mr. Whitman said.
It's a long way from the days of talent scouts combing smoky bars for
the next big sound.
But computer analysis of songs is not necessarily new. A wide variety
of companies spend hours in laboratories breaking down hit songs so
the music industry can stay one step ahead of the market.
The goal is to pinpoint trends in pitch, rhythm and cadence that are
driving consumer spending habits. However, the MIT researchers believe
they've taken the science to another level.
"Some people really care about instrument sounds and complexity of the
music," Mr. Whitman said. "But the 14-year-old teenage girl could care
less, as long as her friends are listening to it."
The MIT method, developed at the school's renowned Media Laboratory,
also takes into account social responses to hit music that are fed
into the algorithms.
The researchers pull data from weblogs, chat rooms and music reviews
-- anywhere a song is being discussed -- and feed it into the
computer, which allows the software to gauge the popularity of a
certain sound.
Once all the information is tabulated, the computer can listen to an
entirely new album and predict how people will respond based on what
it knows about the latest reactions to the music it has already heard.
If it sounds far-fetched, consider this: the system has been
predicting Billboard hits with surprising accuracy over the past
several months. While people may think their musical tastes are
unpredictable and whimsical, they are actually quite traceable, Mr.
Whitman says.
The researchers' goal is to revolutionize the tracking techniques used
by companies such as Amazon.com and Apple Computer Inc.'s iTunes music
store. Those companies compare similarities between songs, add in the
buying history of consumers, then recommend albums that each person
should buy.
Mr. Whitman and Mr. Jehan, who are both musicians, scoff at those methods.
"They say you bought this so you'll like this. But it's really bad for
music because it can only recommend stuff that people have bought a
lot of," he said.
Still, the music industry has been trying for decades to come up with
a reliable system. The standard practice today is to crunch data from
focus groups across a broad spectrum of tastes, which gives hints of a
song's true potential in the market.
New York-based HitPredictor has built its business crunching weekly
data from focus groups, and many of the play lists heard on North
American radio are influenced by the company.
HitPredictor polls thousands of listeners each week on songs that have
not yet been released, then makes prognostications on how the single
will perform.
The company established its credibility in 2002 when RCA used its
method to determine the order in which the singles from Christina
Aguilera's album Stripped should be released to maximize record sales.
Since then, other labels have turned into regular customers.
After crunching feedback data on the Aguilera album, HitPredictor
realized RCA needed to rethink the release order because the focus
groups were unexpectedly reacting favourably to some songs, but not
others. Each prediction the company made in terms of how well each
single would sell eventually proved true in the market.
"A lot of labels put music through our research to confirm their
instincts," said Doug Ford, co-founder of HitPredictor. "They've got a
few guys in management that like this song, but marketing likes that
song, so they go through us."
HitPredictor struck gold again in late 2003, when its computers
flagged a blip in the focus-group data. Listeners, who are fed random
songs and asked to rate them, were repeatedly highlighting a
little-known U.S. band called Crossfade, which the big labels had
passed over.
Mr. Ford went to Sony Music and told the company to consider listening
to HitPredictor's computers and focus groups rather than their own
talent scouts. In 2004, the band sold more than a million albums.
Despite the performance of HitPredictor, the researchers at MIT aren't
looking to build another software program that simply picks commercial
hits. Mr. Whitman and Mr. Jehan's goal is to expose the world to a
wider variety of music.
Forecasting what songs people will like before they hear them is easy,
they say. In many ways, it's been done for decades without computers.
Finding good songs is much harder.
"There's too much music out there and its really hard to figure out
what you want to hear," Mr. Whitman said.
"So we have systems here that are automatically identifying what
people like, without knowing much about them . . . we're trying to get
in between the audio and the audience."
--
"Facts are meaningless. You could use facts to prove anything that's even
remotely true!" -- Homer J. Simpson
