Slashdot Mirror
Using Fractals To Classify Music
Brian McLaughlin writes "A company is working on software that can classify music with fractals and make it easier to find the tunes you want on the Web. Apparently, one can detect the type of music (jazz, heavy-metal, in-between, etc.) by detecting fractal patterns in the music."." I'm looking forward to the day when my music can be indexed and crossreferenced every which way: artist, tempo, year, style, similiarity, heck I wanna know when the Beastie Boys sample The Beatles and be notified and give the option to follow up on the samples within the songs. Someday... I hope.
What happens if you feed that program disimilar styles of music, say Metalica, Portishead, and Beatovens 9th?
It fails to find any kind of meaningful consistency, the ATN's knowledge base becomes underpopulated, and the final product is utterly bland and devoid of content. (Wow, so that's how they compose new songs for Britney Spears records!)
By the way... Beatovens? That's a damned cool name for a band! I've got dibs on it!
Also, if I feed the same songs in again do I get the same song out the system a second time, or is it just the same song as the first time?
No, the generative part of the process is randomised, so you merely get a different song in the same style. Look at the example MIDIs in the EMI web pages; there are a handful of generated Nocturnes in there, IIRC, and they're all different.
To the editors: your English is as bad as your Perl. Please go back to grade school.
Such Classification could also be used to detect the type of music you want to hear on a mix. For example:
Let's say you make yourself a bunch of mix CDs and stick them into your 200-CD tray (it makes more sense to have an MP3 player, but hey...), and from those CDs you only want to hear techno, or only want classical because your parents are coming for dinner. The auto-detection functions allow for limitless possibilities for music playback.
Maybe this could also be used on TV to filter out stuff you don't want to see? Like a quick auto-seek for a channel that doesn't have a cxommercial running?!
Couldn't the reverse, then be done, to take a fractal "fingerprint" of a type of music (say, jazz), add some variables and come up with original music?
It's 10 PM. Do you know if you're un-American?
Actually, this could be used to shoot down the popular "filename says sandman, but is not metallica" arguments. When Metallica told Napster to ban 300K users, it was because of a filename which does not infringe on copyright. Now a band can say, "Hey, not only does it have a suspicious filename, but it follows the pattern of our hit song exactly.
This is cool technology though...maybe someday we will be able to prove for certain that Vanilla Ice did use David Bowie and Queen's 'Under Pressure' in 'Ice Ice Baby'. Hmmm
It's actually not that difficult. The algorithim (that I know) is a fairly straightforward neural network. I took an Intro to AI class while I was in college - had a friend who's final project was a system that could determine the type of music on a CD that was currently in his CDROM drive.
He trained it first, with 2 CDs from each genre from his collection. He then went through and had the system analyze and 'guess' the type of music. From his collection of ~100 CDs, it correctly identified the different types about 95% of the time.
Now, I don't know the constraints that these folks are putting on their software. But if Erik could code up a working model for a 200 level AI class, I'd hope that this company can handle details.
JJ
My favourite music is Einstuerzende Neubauten, and anything else just isn't noise.
Any music which you are not familiar with will sound "all the same" until you are familiar with it. People have said that about jazz, pop, classical, heavy metal, techno, hip hop, punk - anything in existence. If you think any genre "sounds all the same", I submit that you are not familiar enough with it to be able to understand what the different artists are trying to achieve. I would say that you need to be familiar with at least 100 recordings of a genre, and have read several books about it, before you are qualified to even begin considering judging it.
...) are no different.
Something like country is tremendously diverse, and is also one of the oldest recorded musics. Few music lovers don't love older country to begin with, and when you add things like alt.country and bluegrass and contry-folk (each of which have a bunch of different sub-genres), you have a tremedously respectable music, and inarguarbly one of American's two or three finest traditions.
Of course, if all you've heard is Shania Twain and whatever else they play on the radio, you've missed out. Like any genre, the best country music is not played on the radio. Judging country music by Garth Brooks makes about as much sense as judging jazz by hearing only Kenny G, metal by only Bon Jovi, rap by only Snoop Dogg, and classical only by Charlotte Church. For any genre you need to dig deeper than the tunes played on the radio, and country (and hip-hop, and jazz, and
I was skeptical but it actually works quite well. Even better, they're an open source shop through and through: Apache, mod_perl, mysql, etc.
Using fractals to find simalarities in music should allow us to prove that there is only one rap song and that it is just being re-recorded by everyone with a microphone. Same for most classical, country and pop.
I recall there being a high-end, all-analog radio tuner that used a special wide-screen, green phospor, cathode-ray tube to display the entire FM radio spectrum at once. The nice part is, it allowed you to tune to the center of what was being broadcast. (Those big transmitters did drift.) And it totally fit with the analog-only mindset of being forced to listen to a digitally/decimally perfect frequency.
The long-time users of those systems said they could tell what type of music the station was playing by the frequency distribution, and frequency energies being used. Some said, for their favorite station, they could even tell what period of music was being played, or if it was one of their favorite composers.
If you've seen interviews with many artists, you've probably noticed how most of them hate to be classified. If this really works, it would kill a lot of pretension. I can see it now:
Reporter: Despite the fact that you're considered a rock artist, you seem to be having a great deal of success amongst the country and even R+B fans. Why do you think that is?
Artist: Well I don't really think of myself as a rcok musician. People are always trying to classify my music as pop, or hard rock, or soft rock, or whatever. But I don't restrict myself to those terms, I just think of myself as an artist, and I think my work really defies being simply classified as rock.
[Reporter looks down at a laptop, hits a few keys]
Reporter: No, the computer says it's clearly rock music. Not country, not big-band, not funk, not innovative-genre-transcending-art, just rock music.
Artist: Well, yes, but it's really-
Reporter: That's all the time we have for now. Tune in tomorrow as I interview another popular rock band.
I hope it's better than the BAIR system for recognising naughty pictures (http://slashdot.org/article.pl?sid=00/07/21/1516
</skeptical>
Mike.
Tales from behind the Lagom Curtain
That is the shortest story I think I have ever seen on slashdot!
I can't myself see how much detail can be garnered from the amplitude alone of the notes in a musical piece. I can see how heavy metal and jazz would be quite different, but how about jazz and drum'n'bass, they are incredibly similar forms that would be distiguishable by the underlying beat rythms (or perhaps more likely through the persistance of instruments). Would a fractal based on note amplitude grab this? I can imagine it might, but if you threw it the whole gamut of dance music (acid, trance, acid-trance, garage, girly garage, ambient, techno.....need I continue) I am sure it would fall apart as the large scale use of compression alone would bring these musical forms incredibly close together in terms of amplitude (if everything is at 100db post-compression as often happens with dance music).
Another useful bow in the arrow of anyone interested in categorising music, but I feel that a full quiver of tools is always going to be needed to even come close to trying to do this job.
Never underestimate the dark side of the Source
This could create one hell of a visualization plugin for Winamp...
Sleep: A completely inadequate substitute for caffeine.