Slashdot Mirror
AI Bots Pick The Hits of Tomorrow
Wolverine Inspector writes "The Music Industry uses a product called HSS (Hit Song Science) made by Spain's Polyphonic HMI. According to The Guardian "while no one's talking about it, it seems that the whole record industry is already using AI to choose hits. From unsigned acts dreaming in their garage, to multinationals such as Sony and Universal, everyone is clandestinely using a new and controversial technology to gain an edge on their competitors."
Even though it costs about $5,200 US/$6,500, many artists are starting to buy it to help them write succesfull songs."
...on Slashdot.
*sigh*
This is not AI. The music companies are using clustering technology.
The basic idea is that you measure certain characteristics of a song,
such as voice quality, cadence, etc. I'm sure the actual
characteristics used are much more complicated, but the idea is the
same. Once you have your characteristics you can build a three
dimensional vector out of a song. After you have your three
dimensional vector, you can then use many different algorithms, one
such is the Bi-secting K-means algorithm to group the songs together.
After you have built your cluster, you take a new song, run it through
the process and check to see how close it falls to a "hit" cluster.
We use this same process for document classification at my work, and I
don't think it bears any relation on AI. As I stated above, it's a
rather simple grouping technique.
There is a downside to this technology though. By measuring how close a
song is to previous hits, you are guaranteeing that all new songs will
be similar to old hits. This type of system tends to minimize or
eliminate fresh new types of music.
(why the word wrapping? Emacs auto-fill-mode)
Doug Tolton
"The destruction of a value which is, will not bring value to that which isn't." -John Galt
This was discussed last November, which was a repeat of the same tech from February.
A quick search for "polyphonic" in the music category would've easily picked this up, they're the only 3 matches!
Oh really? Because many people are composing actual music today right? ...
Britney Spears, et al != Current Music
There is other music out there.
Does this remind anyone of the Monty Python skit where they use mathematicians to create the world's funniest joke, and use it to get Nazis to die laughing?
Actually, there was a joke writer who came upon it by himself and died. There were some scientists later in the sketch, but that was later when they were in isolation to translate the joke.
From what I know about DJ's "freedom" to pick songs, a certain number of songs played during their shift (typically four to six hours) must be from the approved playlist. Depending on the location of that station (and therefore how important the market is), they might have to play more of the "required" playlist or less. (I seriously doubt that the stations here in Columbia, SC are held to the same requirements as a much more competitive area like NYC.)
Usually, these requirements are structured so that a DJ can't play all of the "required" songs in the first hour or so of their shift and then play anything they want for the remaining 3-5 hours. (More's the shame.)
From what I can tell of the local stations, it seems to be about 75% of the songs they play are from the required list and the rest is up to the individual DJs.
Kierthos
Mr. Hu is not a ninja.
Does this remind anyone of the Monty Python skit where they use mathematicians to create the world's funniest joke,
o ke.html
"This man is Ernest Scribbler... writer of jokes. In a few moments, he will have written the funniest joke in the world... and, as a consequence, he will die... laughing."
http://www.jumpstation.ca/recroom/comedy/python/j
... are intricately related. Many AI techniques are forms of statistical inference or statistical classification techniques. Some neural nets implement grouping techniques not that different from k-means.
Any box which learns from a set of data in order to predict future data by implicitly extracting trends and patterns from that data is an implementation of some form of statistical inference algorithm and is subject to all of the general results statistics has to offer about such algorithms. Conversely, statistical inference algorithms are often implemented in ways associated with AI, for example as neural nets.
Given this situation, it's hard to define the boundaries that separate artificial intelligence, pattern recognition, statistical inference and classification and the rest. Of course, there is a legitimate question as to whether such techniques actually mimic genuine intelligence even in principle, and there are other approaches.
From the point of view of terminology, there is a huge range of techniques that can be called AI, and statistical inference is one of them. If you call a VLSI neural network implementing a statistical inference algorithm "AI", then why not call a normal computer implementing a statistical inference algorithm "AI"? Besides, AI sounds a hell of a lot sexier than statistics when you're trying to extract maximum dough from the ample coffers of the recording industry.
"The Milliard Gargantubrain? A mere abacus - mention it not."
While I agree with your sentiment, just for clarity's sake I have to point out that Maroon5 has been around for quite a few years, previously by the name of Kara's Flowers.
do not read this line twice.
The differences were consistent. It was obvious that mainstream versions had african musical characteristics (rhythm based) whereas the popular versions were more european influenced (melody and harmony).
If you listen to what is mainstream music today, the same patterns emerge. Virtually all pop songs follow the same template. The chorus and verses are always in the same places, the breaks are always at the 3/4 mark etc...
The beats are also important. Pop music relies heavily on the 4/4 beat, with the accent on the downbeat. African influenced musics have a lot of syncopation (accent on the off beat). Syncopation is what makes something "funky".
Lastly, there is a great book called "How to Have a Number One The Easy Way" by the KLF. Its online here: http://www.tomrobinson.com/work/klf.htm
Just follow this to the T and they guarantee you a hit. Its really just a matter of following certain rules and watering down to the least common denominator.
Otherwise, you could put a genetic algorithm and a synthesizer on the job. Use the HSS application as an evaluation function, and let it crank until it had composed an optimal song. Or just run every free MP3 on the web through. (Now that would be a good idea. Somewhere, there may be a garage band that doesn't suck.)
There's a similar program to predict Wine Advisor scores. If that were easily available, people would be synthesizing the optimal wine.
Seriously though, all 'western' music is based on the same set of notes arranged in a variety differnet scales and/or chords, and played with one of a variety of progressions. What differntiates between genres, or what is good or bad is a matter of what type of scale used and what chord progressions are used.
You want to write a blues song? Most of the time your going to use a pentatonic scale with a I-IV-V chord progression. Sometimes the progression will repeat every 8 bars, sometimes after 12 bars. Sound formulaic? It is, but there is room for expression and improvisation within that framework, but it is that combination of scale with progression that tells your ears, weather you know it or not, that it is a blues song.
In the article they say that U2 maps to the same cluster as Beethoven, and that Van Halen looks similar to Vanessa Carlton.
Since this shows that 'hits' in dissimilar styles of music can map to the same space on their model. There is the possiblity that a piece of music done in an up and coming genre that maps well could end up being produced where previously it wouldn't have. (Don't bet on it though).
I want to shoot the messenger!
In the field of machine learning, it's considered a major no-no to quote performance figures based on your training data.
The typical way to validate your results is called an N-way cross-validation. You split the data into N parts, and perform N training runs. Each run uses N-1 chunks to train, and tests on the remaining chunk. Then you average the results to get a general performance estimate, or you can use a T-test to compare the results against another algorithm.
This report would have been rejected immediately from any academic journal of any significance. It's a fucking joke.