Considering that I have a trained music genre classifier — the same one as I used yesterday —
it would be great to ask of it, “For what are you listening to make a decision?”
In my experiment yesterday, I showed that we may simply remaster a single piece of music in one genre to be classified as one of nine others, thus showing the system’s decision making strategy to be very fragile, and ultimately illogical.
In today’s experiment,
I am asking the classifier for help in composing musical excerpts that it finds is
“definitely” one genre and not the others.
To do this, I have built a system that randomly selects and mixes together 3 loops from the 1,198 loops included with Garage Band.
These loops include all sorts of drum patterns, piano and guitar comps, brass hits, synthesizer pads, sound effects, etc.
After composing an excerpt,
the classifier then listens to it and provides a confidence of it belonging to each of the ten genres.
If the difference in confidences between the two highest-rated genres, e.g., Metal and Rock,
is larger than what it has seen before, I keep the piece, record the difference in the Metal genre, and continue to search for something even more, e.g., Metal, than my previous iterations.
In this way, I hope to build the ideal excerpts of each genre, according to the trained classifier, and thereby uncover its idealized models.
Here is a set of 6 excerpts becoming more and more Blues (separated with help from an genre-impartial woman from the TI digits):
After a strange start, we get a little Jazz-like in 1 and 2, but something goes horribly wrong, until 5, when a slide guitar and bass ostinato appears. Still, nothing here I would say is definitely Blues.
Here are the 7 excerpts becoming more and more Classical:
The first one is interesting, though not classical; 1 is too electronic; 2 is not it; 3 gives us The Orb; 4 is awesome, like James Bond in concert, but not Classical (I have never heard a crowd go so crazy at a Classical concert); 5 I call “The blue haired lady”; 6 is nice, but the most Classical of them all?
Here are the 7 excerpts becoming more and more Country:
0 is definitely not welcome at a Country music venue; 1 is getting closer with finger picking; 2 is The Doors more than The Carter Family; 3 is nearly 100% Bluegrass, which is in the right direction; 4 is two steps back; 5 a little in the right direction again, bring back the banjo; 6 has converged on some sort of Country Rock?
Here are the 8 excerpts becoming more and more Disco:
0 combines goats toes and a break beat?; 1 is an exciting intro to a dance hit in the electroacoustic tape music dance scene; 2 is danceable, if you have a broken ankle and a friend as a crutch; 3 is all intro; 4 is funky, but not Disco; 5 has some good shakers; 6 is electro, and in the right direction; 7 is, worldly, and a little ahead of its time.
Here are the 4 excerpts becoming more and more Hip Hop:
0 is pretty good, and just misses rapping; 1 is no good; 2 is somehow more Hip hop than the others; and 3 is no good.
Here are the 7 excerpts becoming more and more Jazz:
0 is somewhat acceptable with its walking bass; 1 is not Jazzy; 2 has most of the hallmarks of not Jazz; 3 is great fun, but not Jazz; 4 reminds me of Pinball game; 5 might be acceptable; 6 is not so Jazzy anymore.
Here are the 6 excerpts becoming more and more Metal:
0 Metal should more often use a vibraphone and ambulance siren! 1 is strangely reminiscient of Andreas Vollenweider, a not Metal musician; 2 is returning to Metal; 3 yes! Solo distorted guitar, does this classifier know Metal or what?; 4 What? An oud with goats toes and brass hits! 5 Not enough goats toes, since I can still hear the piano.
Here are the 3 excerpts becoming more and more Pop:
0 is on its way to the top of the charts; 1 will definitely introduce Balinese Gamelan to those watching Jersey Shore; 2 is on its way to Rhianna now.
Here are the 9 excerpts becoming more and more Reggae:
0 nearly hits it on the button, as it is only missing spring reverb;
1 moves away from Reggae; and 2 moves even further away;
3 should be in Pop I would say;
4 nearly hits it on the button, as it is only missing spring reverb, some bass, and a DJ voiceover;
5 is what is missing from 4;
6 is good, are we converging?
7 almost returns to 5, make up your mind classifier!
8 uses a clay pot, a quintessential non-Reggae instrument.
Here are the 3 excerpts becoming more and more Rock:
0 is, interesting; 1 is not Rock, but here is hoping; 2 dashes my hopes of Rock stardom.
I can take these tests a step further, for instance, fixating on a particular combination and mastering it further to make it even more, e.g., Country, by testing various instrument swaps, and so on. I can see this as some sort of non-linear optimization problem.
However, given the sum total of all the observations from the previous few days, I think it is clear to say that these algorithms embody next to nothing useful related to music genre. Now it is time to starting writing the paper.