Bach is not Grunge
We are building a system that is trained to predict social
tags for music. Our system will try to 'learn' what social tags
should be applied to an mp3 file based solely on the audio
content. The system works rather well for most kinds of music, but
we've been seeing some particular problems with music that is labeled
'grunge'. Bands that you'd think would be labeled as 'strong
grunge' - bands like Nirvana and Pearl Jam - are only mildly grunge
according to our system. This has been a bit of a mystery, until
this week when we added a feature that lets us search through our
predicted tag space - with this feature we can find the most 'metal'
(Good Riddance, Atheist), the most 'folk' (Bob Dylan, Cat Power), and
the most 'electronica' (aphex twin, Jake Fairley). However for
the most 'grunge' we didn't see Nirvana, or Pearl Jam (nor even Courtney
Love), instead the most 'grunge' artists were Johann Sebastian Bach,
Charlie Parker, Glenn Gould and Wolfgang Amadeus Mozart. Clearly
this isn't right - so we dug in to see what was happening. We
looked at the training data for the 'grunge' tag - sure enough it was
filled with non-grunge artists, classical, jazz and spoken word - so
instead of learning 'grunge' we were learning randomness (and
woe). Thierry quickly found and fixed the problem (it was a
problem with how we were normalizing tag counts) - so now we are
retraining a bunch of our tags - with hopefully even better results
coming. Clearly though, we need to think a bit more about adding
some basic sanity checks - so we don't happily go though life thinking
Bach is grunge.
I'd love to see a break through in that field! :-)
Posted by elias on August 21, 2007 at 05:56 PM EDT #
Will your tagging algorithm ever be able to tell the difference between impressionism (Debussy), serialism (Boulez), expressionism (Varese), neoclassicism (Prokofiev), minimalism (Glass), drone (La Monte Young), nationalism (Sibelius), romaticism (Tchaikovsky), baroque (Vivaldi), sturm und drang (Haydn), or lyricism (Schubert)?
All classical music can't be lumped together into one bundle. (Someone once told me that the main differentiator between classical and pop is: anything longer that 3 1/2 minutes is classical).
There are a zillion styles in so-called classical music. And a zillion composers around the world, very much alive, writing music for "classical" ensembles and audiences.
Just listen to one of my radio programs (rchrd.com/mfom) and try to figure out how to categorize the style, or even the period.
Any reasonable ear can tell the difference between Schubert and Debussy. They use the same 12 notes per octave.
But it's the difference that is amazing.
Is anybody looking into this? Or is just pop music of interest (because of the inherent commercial value .. one thing that so-called classical music lacks -- commercial value)?
I'd like to know.
Posted by richard friedman on August 22, 2007 at 03:24 AM EDT #
Richard:
I don't know how well we'd fare distinguishing between different periods and styles of classical music - we haven't tried it yet. To train our system, we need lots of well labeled music. We can get this data via web mining music sites like last.fm - but this does skew our models to be primarily focused on popular/commercial music. There just isn't a good source of social tags for classical music. If there was such a source, we would use it. Classical music fans don't seem to be as aggressive about tagging as indie music fans are.
Classical music is not forgotten though. For instance, at this year's MIREX (the Music Information Retrieval Evaluation eXchange), there is a 'classical composer identification' contest. The task is to identify the composer of a number of pieces of classical music from the baroque, classical and romantic periods. The identification is to be made using only the audio. The contest will include music by Bach, Beethoven, Brahms, Chopin, Dvorak, Handel, Hayden, Mendelssohn, Mozart, Schubert and Vivaldi.
You can read more about the evaluation here: http://www.music-ir.org/mirex2007/index.php/Audio_Composer_Identification
Thanks for the comments!
Paul
Posted by Paul on August 22, 2007 at 06:36 AM EDT #
I wonder if Bird and Bach use the same quietLOUDquiet structure that Nirvana borrowed from the Pixies? If that kind of structural element is used as an identifier in the acoustic analysis, I wonder if that would be seen as a similarity.
Posted by Zac on August 24, 2007 at 02:47 PM EDT #