## Every Noise At Once

Glenn McDonald is the guy who wrote the amazing, obsessive, beautiful music blog The War Against Silence, now mostly dormant.  I admire him for writing tens of thousands of words about Alanis Morissette, whom he, and I, and maybe nobody else, still consider an important cultural figure.  He’s also a pretty hardcore data analyst.  I’ve often fallen down the rabbit hole of his analysis of the Pazz and Jop ballots.

Now he works for Echo Nest, the Greater Boston music startup that sponsored the Music Hack Day I participated in a couple of years ago.  And his latest project, Every Noise At Once, is a map of all music.  Seriously!  A map of all music!  By which I mean: an embedding of the set of genres tracked by EN into the Euclidean plane, and, for each genre, an embedding of bands tagged in that genre.

## 33 thoughts on “Every Noise At Once”

1. DS says:

I didn’t last through the whole Japanoise clip.

2. JSE says:

You used to be cool, DS!

PS: If you’re going to call me obsessive, which I don’t dispute, then I guess I’m sort of obliged to point out that you spelled my name wrong…

4. JSE says:

Well, crap, I got it wrong on FB, G+, and Twitter, too, then. I even thought to myself, “You know, I don’t actually know how to spell his name, but I’m just going to go with my instinct.” I managed to get both n’s at least!

Now that you’re here, any chance you’ll tell us more about what’s going on underneath? Do you have a similarity metric and then do multidimensional scaling, or is this some kind of PCA we’re looking at, or what? Or is this proprietary?

I suppose it’s a nice exercise in data analysis, but am I alone in thinking this is a pretty awful way to think about music? I’m not feeling terribly articulate about it at the moment, but this kind of micro-classification that a certain kind of music geek is attracted to really gets me down. Maybe it’s that it makes a claim of universality (“a map of all music”) that it can’t possibly back up, and then it’s just laughably myopic. To pick the low-hanging fruit, this chart has 50 of kinds of house music over on the left, and then “classical” up in the upper right. (Really he should have left the classical corner out rather than trying to fit it into his scheme.) So the lack of perspective is one thing.

But more than that it’s the idea that trying to group songs and bands into these tiny groups that can’t have more than a couple members is a good way to talk about music. Why not just say, this band’s sound is like that band’s in this way, and like that band’s in this other way? I suppose someone who loves this kind of classification will say that’s all they’re doing, but there’s something about encoding it into phrases like “vapor wave” that I find really annoying.

Sorry to ramble. As I said I’m not feeling very articulate about it.

If he made the same chart with individual songs or bands, scoring a few characteristics and projecting them down to two dimensions, it might be more interesting. Maybe.

By the way, I have the same gripe with a certain kind of beer geek.

6. JSE says:

“A map of all music” is my phrase, don’t blame Glenn for it! Of course it is overstated.

The thing is, I think that whatever genre you really focus on, the microgenres within it actually do seem important to delineate. I think of “twee” as a thing I care about it (and indeed, bristled that “Club 8” is classified as “twee” on the graph when I think it should clearly be “Swedish indie pop” — they are not that twee!) but I sort of chuckle at “symphonic black metal.” But if I actually listened to black metal it would be just the reverse! So I think the right stance is to accept all the fine distinctions as legitimate and not a waste of time.

The unevenness of the granularity is half my gripe. But I don’t think it’s a failing of this chart especially, I think it’s a failing of the whole philosophy that produced this chart. Someone who’s drawn to this kind of classification has a reasonable way of thinking about the music they know well, and they just can’t resist analytically continuing it (sorry) to a much wider swath of music where it becomes silly.

So what does it mean to say that Club 8 is Swedish indie pop rather than twee? Because I don’t doubt that there’s something worthwhile you’re trying to say by saying that, but I don’t think I’d want to say it that way.

8. In any metric which distinguished between harmonic structures, I would imagine that the distance between Bach and Rachmaninoff would be some large fraction (say bigger than a half) of the diameter of the entire graph, but here they are immediately adjacent, so something is amiss.

9. The layout is not done with similarity, it starts as a straight XY scatter-plot and then I have code to bump the names around so they’re all readable. Internally I have a tool that can do XY plots with any of a dozen or so metrics on each axis (and for label scaling/weight). The pair I use for these is a measurement of electric (left) vs acoustic (right) on the X axis, and a harder-to-describe measure of sonic density/uniformity on the Y axis. The adjustments for readability are all on the Y axis, though, so the deviations from the data are bigger as you go down.

On the main map, those metrics are averaged over a few thousand songs for each genre, so there’s a fair amount of mass behind them. On the individual genre maps, they’re averaged over 20-ish nominally genre-appropriate songs for each artist, so there’s more random variation at work.

As for genre distinctions, philosophically, I like them. Obviously, or I wouldn’t be doing this. I don’t consider it harmful to label things, at least when the labeling is non-exclusive. (I agree that trying to jam a band into one and only one genre is a waste of time.) Think of it like this: you like what you like, and if we can find a label for some bits of what you like, you might be able to follow that label to more stuff you like. You can do that with individual artists, too (the “similar artists” thing on Rdio is driven by the same Echo Nest data that ultimately feeds into these genres), but it takes a lot longer and it’s easier to get lost.

If you want to play with a similar thing that scatter-plots individual songs by a given artist, go here: http://static.echonest.com/ArtistX/. That one was done by my Echo Nest colleague Paul Lamere, and uses a subset of the audio-analysis data that also feeds into this genre thing.

Also, Nick, you say “trying to group songs and bands into these tiny groups that can’t have more than a couple members”, but in fact all of these genres in my map have at least hundreds of arguably relevant bands. There do exist “genres” that have only a handful of bands “in” them, but I don’t use those. That’s been the most amazing thing to me as I’ve worked on this: niches that sound at first like somebody made them up to describe their own one band turn out to be little subcultures that keep going as far as you look.

And yeah, as a listener I think the more distictions the better. I like gothic symphonic metal better than gothic metal or symphonic metal, and uplifting trance better than progressive trance. And when those are labeled things we can both experience and talk about, we get a vocabulary.

10. And yes, there are other metrics that distinguish between Bach and Rachmaninoff. The two I’m using (uniformly) for the artist-maps don’t. Neither are harmonic in nature. I could do classical with different axes, but I thought it was more interesting (and less confusing) to use consistent axes for all the genres.

11. Frank says:

Just being able to see two hundred names of brutal death metal bands, all in one place, was by itself worth the price of admission.

12. Also, in all cases the list of artists for a genre is just an attempt at finding the most definitive and illustrative examples. Absence is not intended to imply that a band is “not” that genre, just that our data suggested other bands were better fits. And the band lists cut off arbitrarily somewhere between 200 and 250 for each genre, depending on exactly how the math goes. So we have Club 8 as band #167 for twee pop, and they don’t quite make the 250 for swedish indie pop. They don’t make the list for indie pop, either, where they also arguably apply. But clearly in many of these genres, out in the world, there are way more than 250 bands. How many “rock” bands are there? Lots. But if I put 4000 bands on a page, it would be unwieldy. Likewise, across there whole set here there are close to 50,000 bands, and if I tried to put all of those into one map, as Nick wondered/suggested, it would be a useless mess.

So honestly I don’t believe I’ve misapplied a method of analysis from one place to another where it “becomes silly”. You’re free to think it’s silly, period, but I think it’s not especially more or less silly in different parts of the space due to the nature of the music.

13. davidzb says:

Obligatory (though not as awesome) link, especially after Frank’s comment: http://mapofmetal.com

> niches that sound at first like somebody made them up to describe their own one band turn out to be little subcultures that keep going as far as you look.

> And when those are labeled things we can both experience and talk about, we get a vocabulary.

Of course, you need a language to be able to talk about music, and this is how gothic symphonic metal fans talk if they want to be understood by other fans, so I should get over myself. But… I feel like microgenrefying distorts the discussion somehow – I don’t know, it’s running away from me as I try to pin it down, but it just seems deeply wrong-headed. I’ll keep thinking.

What do you like about gothic symphonic metal better than the other two?

> I think it’s not especially more or less silly in different parts of the space due to the nature of the music.

What do you mean? You can’t possibly claim that the reasonable-looking atlas of metal genres in the upper left and the “indian classical/show tunes/chinese traditional” area over on the right are equally meaningful.

16. There are denser and sparser areas, sure. If I had more subvariations on indian classical music, which I’m sure exist, then it would look less lonely. The set of genres is a work in progress. I just added “chinese traditional” an hour ago. Clearly there are more kinds of classical music than just “classical”, too. But I don’t think that makes the metrics less applicable.

It’s worth repeating that this isn’t based on similarity. There’s nothing stipulating that the metal styles should be in a cluster, they’re in a cluster because they share aggregate audio qualities. Likewise, I’m not stipulating that indian classical music is like show tunes, I’m discovering that those two culturally unrelated forms are similar on at least these two audio axes.

17. Richard Séguin says:

“harder-to-describe measure of sonic density/uniformity on the Y axis”. As is, this definition is not particularly helpful.

The word “classical” is often broadly applied to everything from early music to twelve tone music (you know—anything having to do with violins, lutes, pianos, and such, wink wink). But “classical” is also used to refer to a specific time-period/genre of music, and in this case classical would not include, for example, early, baroque or romantic music.

18. Up is more sonically dense with consistent volume levels, down is more sonically spiky. E.g. the blurry distorted wash of black metal at the top, the sharp choppiness of reggae at the bottom.

And yes, the current “classical” is a catch-all in the same way that it would be as a bin label in a record store. If we still had record stores.

19. Echoing a previous point, the the axes you choose to project on may well preserve the metric better of your data in certain neighbourhoods rather than in others. Naturally, your choices of projection are affected by your own particular interests and tastes. Here’s a suggestion: suppose the initial set is given by some $\Sigma \subset \mathrm{R}^N$, and your graph arises from considering some projection $\pi: \Sigma \subset \mathrm{R}^N \rightarrow \mathrm{R}^2$. Then, for a point $x \in \Sigma$, there may be a sensible way of measuring the extent to which $\pi$ preserves the metric structure on $\Sigma$. (I have various naive guesses for such a measure, but someone who thinks about data would have much more sophisticated ideas than I can offer.) In this way, you would have a measure at each point in your image describing how well that projection reflects reality, and you could color code the points according to the values of this function. For example, in the current graph, it might be instructive to have certain points of the graph in some light colour (yellow, orange) indicating that the projection is giving a good measure of similarity, whereas within classical many points would be the opposite (say purple), indicating the large loss of data. That would help clarify which proximities are artifacts of the axes you have chosen (which, as you say, still conveys some information) and which are reflective of deeper similarities.

20. gowers says:

If you ever do try to create a map based on similarity, I’d love to see it …

21. @gowers
>If you ever do try to create a map based on similarity, I’d love to see it …

Me too. But I think the concept of similarity of music is subjective. Typically you can appreciate much finer differences between musical pieces of your favorite kind while if you haven’t been exposed to a certain genre at all, everything of that particular genre sounds like pretty much the same to you. It’s a little bit like how a person who grew up in a place where one ethnic/racial group totally dominates the population tends to feel certain other groups of people resemble to each other when it’s just an illusion caused by their unfamiliarity with certain facial features.

I’m very interested to know what kind of scientific measure experts would use and how they define similarity between songs, musical works, etc. when creating a map for laymen. But I’m guessing a scientifically rigorous similarity map of music would look reasonable only to computers or someone who knows every sort of music in existence very well. And if a map looks fair and reasonable to me, a person with a totally different cultural and musical background may say it’s a heavily biased piece of poop. It’s still quite intriguing though if the science behind it is sound.

Anyway, since math is also a theme of this blog, here’s a map of scientific journals based on citations:

http://www.vosviewer.com/maps/

> The set of genres is a work in progress.

This sounds like “If only I knew all the microgenres of classical music and had examples of each then that part of the atlas would be good.” But people who listen to classical music don’t think about it that way – they think about composers, and genres in the sense of string quartets, piano sonatas, lieder, etc., and relatively long periods (e.g. late romanticism). My point is that your model of organizing music isn’t valid for music that’s far away from what you know well. I don’t know anything about Chinese traditional music, or Irish folk music, but I doubt your model is valid there either. Your atlas would be better if you left them out.

From what I know about Indian classical music, though, you’d love it – classification-wise, those guys can go toe-to-toe with the most pedantic techno fan.

23. It’s not clear to me how I would calculate or represent “loss of data” intelligibly. I think what you want is a sense of where the juxtapositions on this 2D map represent more or less adjacency in the full space (which is currently 8D), but that’s a function of pairs, not “the space”. I can imagine a network diagram with weighted edges between every pair of nodes, but I’ll be honest that I haven’t the slightest interest in generating or looking at such a thing. I think these two dimensions are OK for practical exploratory purposes as is.

24. Also, I added baroque, classical period and romantic. Seems to me that they work well enough.

25. JSE says:

26. Surely representing “loss of data” intelligibly is the sine qua non of representing high dimensional data sets via projections, but I understand your interest lies in a different direction, fair enough. Saying that “…that they work well enough,” however, is just a reflection of your discernment (or lack of it) when it comes to classical music. I mean you could put 90% of your data points into a single one labelled “rubbish” and that would seem to work well enough for me, but I wouldn’t pretend that would be good enough for anyone else. It’s almost as though you’ve noticed how the frequency of a particular set of English prepositions allows one to distinguish between Dickens, Trollope, and Thackeray, and then claimed that it makes equal sense to apply the same analysis to French literature.

27. If there is music that moves air without the use of either electricity or physical objects, and which does not consist of noises that either do or don’t have drop-offs in volume between them, then yes, I would agree that this current treatment is likely to represent them poorly.

If your English/French literary analogy held here, then we should expect the classical subgenres I’ve added (baroque, classical period, romantic and serialism (in addition to the existing ones I already had like renaissance, early music, modern classical, avant garde, minimal, music concrete, opera, choral and new age)) to be distributed around the map in an unintelligible way.

However, this is not the case. The classical subgenres are close to each other, and far away from music that I think we could agree is very different (metal in one corner, hip hop and reggae in another), and the other “non-classical” genres that are closest to the classical ones are ones with obvious similarities.

So I think it’s pretty clearly and demonstrably working.

28. If my English/French literary analogue held, then one would expect the opposite, namely, that the entirety of French literature would be lumped together at a single point away from all of English literature (one would presume that the density of English propositions in French literature is fairly uniformly close to zero). And then, *within* French literature, one would expect that the results would be unintelligible. And this seems perfectly consistent with the results in this case.

So, if by “it’s pretty clearly…working” you mean to say that, using some choice of projection, you are reproducing your own myopic view of classical music that it all pretty much sounds the same in comparison to 50 types of house music, then yes, it is working. But, as far as saying anything interesting about distinctions within classical music, I think one could produce an equally accurate graph using wordle.

Or we might expect Dickens and Thackeray to end up several inches apart while Moliere ends up millimeters from Ionescu, which is exactly what we see here. To our eyes it looks a bit like the famous map of the world according to Reagan:
http://landofmaps.com/2013/03/05/the-world-according-to-ronald-reagan/
(I hope you take this as good-natured ribbing.)

Importantly, galoisrepresentations and I aren’t clamoring for a map that represents classical music better; we’re clamoring for a map of metal, house music, rap, rock, and pop to know that it’s just that. More fundamentally I’m clamoring for people not to make maps of music, but I don’t expect to see eye to eye on that one.

OK, I should have shut up several posts ago.

30. Did you listen to the audio for, say, baroque vs serialism? They really don’t sound misguidedly or myopically the same to me.

But whatever. If these maps don’t amuse or help you, move on.

31. Karen Snell says:

I can’t get the audio to work with the ‘Every Noise at Once’ map. I click on things and they pop out with the yellow box, but no sound. Any thoughts as to why I’m having trouble? I’m using Firefox and I also live in Canada – I’ve encountered music listening websites that exclude people in Canada. Would this be the issue?
Thanks for any insight you can provide.

32. JSE says:

I don’t know, but I can say that some commenters on MetaFilter also said they couldn’t get Firefox to play the audio.

33. You need a browser that supports MP3 via HTML5 Audio, which Firefox does not. Chrome and Safari do, and at least the newest IE. There are also links at the bottom of each genre and the main map to corresponding playlists on Rdio, if that’s easier for you.