Automatically predicting hit songs
An article in the Digital Music Weblog describes the Platinum Blue Music Intelligence system that claims to provide a crystal ball for the music industry . According to Platinum Blue they can
Platinum Blue is not the first company to make a system that tries to predict hit songs. Hit Song Science also has a system that can tell you whether a song 'sounds like' a hit song.
Similarly, in the academic world, there have been some investigations along these lines. ISMIR 2005 published "Automatic Prediction of Hit Songs" by Beth Logan. In this paper, Beth describes using Support Vector Machines and boosting classifiers to explore if there is some distinguisable thread connecting hit songs. Her results were 'better than random' and that lyric features were more effective than audio features for predicting hits. She concludes by saying that there is indeed some distinguishable thread connecting hit songs and that one cannot simply dismiss claims by commercial ventures such as 'Hit Song Science" as impossible.
Personally, I think that in today's music world, a 'hit' is much more dependent on the facial features and figure of the performer than the music or lyric content. Even more important (and even less measurable) is the social network that a song percolates through. In the book Six Degrees - The Science of the Connected age, Duncan Watts observes that a new idea (or a new song) is like a seed or trigger. Seeds of change, like their biological counterparts are a dime a dozen. Only one in a million may grow to fruition and
Duncan says that the difference between a hugely successful innovation and an abject failure can be generated entirely through the dynamics of interactions between players who might have nothing to do with its introduction. In a world where individuals make decisions based not only on their own judgements but also on the judgements of others, quality is not enough.
So perhaps the best thing that these automated hit detectors can do is to tell us which songs have no possibility to become a hit. They can be the Simon, Randy and Paula that tell us which songs don't have a chance, the songs that should pack up their things and go back to Smalltown Arkansas.
"detect
and analyze the underlying mathematical patterns contained in a song.
This knowledge, encompassing 30+ parameters including, pitch, melody
and cadence, allows a song to be easily compared to other songs. In
this way, the music market, which has often seemed the purview of the
alchemist, is now open to rational analysis."
Platinum Blue is not the first company to make a system that tries to predict hit songs. Hit Song Science also has a system that can tell you whether a song 'sounds like' a hit song.
Similarly, in the academic world, there have been some investigations along these lines. ISMIR 2005 published "Automatic Prediction of Hit Songs" by Beth Logan. In this paper, Beth describes using Support Vector Machines and boosting classifiers to explore if there is some distinguisable thread connecting hit songs. Her results were 'better than random' and that lyric features were more effective than audio features for predicting hits. She concludes by saying that there is indeed some distinguishable thread connecting hit songs and that one cannot simply dismiss claims by commercial ventures such as 'Hit Song Science" as impossible.
Personally, I think that in today's music world, a 'hit' is much more dependent on the facial features and figure of the performer than the music or lyric content. Even more important (and even less measurable) is the social network that a song percolates through. In the book Six Degrees - The Science of the Connected age, Duncan Watts observes that a new idea (or a new song) is like a seed or trigger. Seeds of change, like their biological counterparts are a dime a dozen. Only one in a million may grow to fruition and
... not because that one bears some special unique quality, but because it lands in the right place.
So it is for social seeds as well. What makes success
difficult to predict is that success has less to do with the particular
vision than with the pattern of interactions into the midsts of which
their pinprick falls. In other words, the success of a song has
less to do with the music, and more to do with the social network that
the song moves through."
Duncan says that the difference between a hugely successful innovation and an abject failure can be generated entirely through the dynamics of interactions between players who might have nothing to do with its introduction. In a world where individuals make decisions based not only on their own judgements but also on the judgements of others, quality is not enough.
So perhaps the best thing that these automated hit detectors can do is to tell us which songs have no possibility to become a hit. They can be the Simon, Randy and Paula that tell us which songs don't have a chance, the songs that should pack up their things and go back to Smalltown Arkansas.