Every Noise at Once is an ongoing attempt at an algorithmically-generated, readability-adjusted scatter-plot of the musical genre-space, based on data tracked and analyzed for 3,294 genres by Spotify as of 2019-07-29. The calibration is fuzzy, but in general down is more organic, up is more mechanical and electric; left is denser and more atmospheric, right is spikier and bouncier.
Lots of other cool Spotify-scraping projects by the author at the bottom.
The webmaster of everynoise.com, Glenn McDonald, works at Spotify as "Data Alchemist". Good article on his work here [0]. Excerpt:
"Subgenres are McDonald's business. By going over listener data and identifying patterns, McDonald and his co-workers can identify clusters of artists who might coalesce into a genre—something he’s been doing since his earliest days at the music-intelligence company The Echo Nest, which Spotify bought in 2014. Today, his work with Spotify's data helps listeners discover artists that may have been hiding in plain sight. McDonald’s “data alchemy” helps populate the Fans Also Like sections of Spotify's artist pages, as well as Daily Mix; it also provides a real-time chronicle of how music is developing and splintering into different styles."
Interesting choice of dimensions... So basically how human and how rhythmically distinct.
Edit: so out of curiosity I looked at the bottom right corner and found "tanci". Odd name. Clicked on it, and it's Chinese spoken word artists. Incidentally I practice Chinese so it's perfect.
IIRC, there are 14 dimensions in total, but it’s impossible to represent all of them on a page. So he went for up-down, left-right, clusters, and colors to represent a subset of them.
I don't get it. I mean, I understand what I'm looking at; a word cloud of music genres that is each linked to a sample. That part I got. But when you label your website 'Every Noise at Once'... I kind of expect to hear multiple (perhaps not 'every') noises at the same time.
White noise is equal energy across all frequencies linearly. Sounds hissy because there are more frequencies included within each musical octave as you go higher.
A more even-sounding spectrum is pink noise which is equal energy per base 2 logarithmic bandwidth. Sounds like a waterfall.
You should look up 'black noise' on the chart. It's music generated using MIDI so that the printed score maximises the quantity of black ink used. The "music" aspect is like listening for the whitespace in a sea of sound.
Also, Georgian Polyphony did not involve strings as a general rule. There were regional exceptions, but early Slavic polyphony was generally a capella.
Spotify, and other commercial online sources, have long failed to get genres like "classical" right. It sounds like they're screwing up "Byzantine" as well, which is sadly unsurprising.
I've noticed miscategorizations in other genres as well.
But It's OK. I think the interesting thing is how the musical snippets are different from each other rather than whether or not they're correct in an absolute sense.
It's a step in the right direction, I think. And it's not surprising that the categorizations and recommendations you get on music streaming services like spotify aren't as ridiculously off the mark as they used to be.
It seems like "Byzantine" is just an unfortunate label, which should be "church music", but if you expand it you will see that byzantine chant is present.
It's still too broad. Byzantine chant is an entire discipline in itself, and so are a bunch of other categories included in "Church Music". There's a category for Anglican choirs, so why not for Russian "Greek" chant or Kievan/Znamenny Polyphony? There's a lot of territory here.
The way this concept would really work, in practice, at a streaming service is as the basis for some kind of cluster analysis that could predict what you might want based on your listening behavior and that of others "like you".
Having people use words to precisely pin-down entire genre's is perhaps near the end of it's usefulness.
Where this site really shines is when you get into a genre that links to a series of linked Spotify playlists. Pretty awesome for exploring Intro, Edge, Pulse, playlists of genres of music you like. It's great for discovery of new artists.
Yeah, I've been searching for a specific genre (what to even call it) for a long time and I feel like this site just gave me a treasure trove of artists I didn't know about.
Filthstep didn't sound anything like what I was expecting. But I didn't really know what to expect. So then I selected some genres for which I could confidently propose an exemplary piece. Ragga jungle: precisely correct. Chicago blues: nailed it. Ragtime: the very archetype. Schranz: maybe that's the intro of a track that eventually meanders into schranz, but the clip on hand is way off. Swedish pop: pretty good. Glitch hop: no way. Gospel singers: US centric, but OK. Nigerian Hip Hop: yeah that's sort of valid in the sense that 2Baba is a Nigerian hip hop producer, but that sample is neither especially Nigerian nor particularly hippy-hoppy. Electro swing: all thumbs up. Bury St Edmunds Indie: A truly sublime blend of Sudbury angst and Lavenham melancholy, virtouosly caramelised in Ipswich je-ne-sais-quoi. (Yes of course I'm joking about that last one; the combination of specific location with vague genre is just absurd.)
This seems like a useful tool for discoveling new sounds, but when it comes to finding out what Polish free jazz really sounds like, I wouldn't trust it one bit.
It's a great music discovery service that I've used several times in the past. I've found some good artists this way, and really wish someone would build something similar for fiction books. The only downside is it's tied exclusively to Spotify.
You're not the only one. I have tastes in music, even strong ones, but they don't tend to follow the way genres seem to be split at all.
Defining a genre by its characteristic instrument set, for instance, doesn't match how I tend to react to things very well, but it's a fairly popular way of separating genres, it seems. (I'm not saying I don't understand the use of that metric, it's very, well, available, in the sense of "availability heuristic". But I do not personally find it all that useful.)
[edit] See my comment below: my joke is about the fact that, depending on the order in which you sum an infinite sequence of waveforms, you can create a sequence that converges to any sound you want [1] (as long as those waveforms together span the full frequency space). Note also that a sum over a truly continuous space of arbitrary waveforms is even more ill-defined.
This would make sense if it were physically possible to play every noise at once. (Where would you place the infinitely many emitters? If they have any displacement at all, then there will be points that would not experience destructive interference at all frequencies.)
Well, for starters, it's physically impossible to have an infinite number of speakers playing an infinite number of waveforms simultaneously, so this silly idea does require mathematical abstraction to be meaningful. That shouldn't be too surprising because there are many places in the physical world where we use infinite series to calculate simple finite physical quantities, e.g. when we integrate to find the area of a region.
My point is that if you really sum every possible waveform, the resulting value may or may not converge depending on the order in which you sum them; in fact, it's a well-known property of such conditionally-convergent series that you can actually get any limiting value you want based on how your order them [0]! (let's ignore the fact that the fourier coefficients can take on a continuous set of values). For example, even if you were only allowed to play a single frequency sound wave sin(x) at volumes that are the inverse of some integer value multiplied by a max volume of 1 (in arbitrary units), you may or may not have them cancel depending on how you group the terms in the sum:
sum = 1*sin(x) + -1*sin(x) + (1/2)*sin(x) + -(1/2)*sin(x) ...
were the ith term in the sequence (starting at i=1) is
a_i = (2/n-1)*sin(x) for odd x
a_i = -(2/n)*sin(x) for even x
This is a conditionally-converging series that will hit all positive and negative harmonic coefficients 1/n and -1/n: the even terms cancel each preceding odd term, and the Nth partial sums therefore alternate between 0 and 2 * sin(x)/(N+1), which itself tends towards zero. But you can group these terms in a different order and get a different limit for the sum; in fact, you can group them to get whatever final value you want!
Now, if you extend this thinking to every frequency of sinusoidal wave, you can start summing every pure tone in arbitrary order to get an arbitrary coefficient for each frequency. By picking your limit for each frequency correctly, you can sum your sine waves in a fourier series [1] to get any song you could ever want! And this is while limiting ourselves to discrete frequencies and alternating harmonic coefficients (since it allows us to take a discrete infinite sum).
So the unexplained punchline to my previous comment is that the problem is ill-defined, or rather, that you can view any song as just a specific ordering of an infinite series of other sounds. (You don't have to use sine waves as your basis, by the way; you can use a bunch of different waveforms that look more like "noise" as long as their combination spans the same infinite-dimensional linear space as pure sine waves; you just end up with different coefficients. For example, in quantum mechanics, you can get a sine wave (momentum eigenstate) by summing energy eigenstates (non-sine waves with a specific form) with the correct coefficients.)
That reasoning allows you to get any final signal you want depending on where you stop in the series, not on how you order the individual coefficients. Addition is still commutative. The parent comment was about every sound _and_ its inverse, which can only ever add up to zero.
(Also, you're missing the frequency components there; your math cannot reproduce any sound at all, it can only reproduce different amplitudes of the same sine wave.)
I'm not missing arbitrary frequency components; please reread the penultimate paragraph, where I mention them explicitly. As I said there, you can extend my argument about a single frequency to include all multiples of that base frequency, and since you can set the coefficient for each frequency arbitrarily based on your ordering, you can set the fourier coefficients arbitrarily and in so doing recover any waveform you want.
Also, your point about commutativity is more subtle than you think; it fails for an infinite sum because you have an infinite space in which to rearrange things. Sure, the terms cancel eventually, but you can keep sticking the negative terms farther and farther back in a pattern so that by the time they've cancelled earlier positive terms, there's already a bunch of new positive terms to take their place. The subtlety comes from the fact that you can keep doing this forever, and you can do it in a way where the sum eventually converges to a specific value.
But don't take my word for it. This is an extremely well-known and basic result in mathematical analysis (the fancy math term for calculus and related topics). Again, see links above, or go straight to a proof [0]. If you want a deeper understanding, check out Rudin's Principle's of Mathematical Analysis [1], which explains this and other fun math stuff very well.
[edit] Just to be crystal clear, the Riemann series theorem does not apply to partial sums, which is what you are saying; if you do an infinite sum on a conditionally convergent series (like the alternating harmonic sum, a variation on which I used in my example), then your final result can literally be any number you want based on how you order the terms in the series. You can set it up so that the infinite sum keeps getting closer an closer to an arbitrary value. If this sounds nonintuitive, it's because infinite phenomena are subtle and nonintuitive!! This is a very cool example of how weird things get once you start dealing with the infinite.
You're right, I missed that your sin(x) example was talking only about a single component.
However, cherry-picking a different reordering for each frequency component before doing an inverse FFT really isn't the same thing as playing all the sounds simultaneously.
Anyway, the thing is, we're not talking about an infinite series. This is a thread about digital audio playback, where both amplitude and phase components (I'm going to assume this site uses some sort of DCT-based codec) are quantized, and hence occupy a finite space. No amount of reordering will change that sum.
Yeah for sure. I mean I was just trying to make a joke about divergent series and how "Every Noise at Once" is a deeply vague statement. But you're right that in the discrete arena it is literally a finite sum that can cancel perfectly (assuming that every sound has exactly one representable "opposite" sound in your storage format).
Nope, it's flipped on the horizontal axis (if you picture it as a graph) so when both are summed you get zero (at least in perfect destructive interference).
I use this site a lot, it is pretty good for finding new music especially new genres of music. I discovered Arab Metal last week which was pretty good.
It's sad that it only has one example for tanci, a genre I discovered due to this list, that aside from wikipedia is barely documented on the internet. After some searching I've managed to recover about 6 hours / 58 tracks of tanci in a spotify playlist -- most of it under the 'various artists' tag, one of the reasons why it was difficult to find any artists, I expect. I'm sure spotify has more of it and that it's uncategorized and badly labelled.
At the bottom of the page it describes this a little better:
>The calibration is fuzzy, but in general down is more organic, up is more mechanical and electric; left is denser and more atmospheric, right is spikier and bouncier.
The artist search is very picky. E.g., "cat ctevens" does not work. "yusuf" does not work. But "yusuf / cat stevens" works.
Some of the genre memberships seem odd. Barry McGuire is bubblegum pop? As is Zager & Evans? Roy Orbison? That whole category seems to be a weird mashup of what I'd expect in bubblegum pop plus a random dump of '60s rock.
For artists that appear in more than one genre, it seems to use the same sample clip for all of them, so don't be put off from checking out an artist in a genre you like because the sample doesn't fit.
The examples this site chooses aren't the greatest... It maps an entire artist to a genre, so when one artist makes multiple genres of music the examples end up being very off. Take dubstep for example, the artist they chose for the example is Skream, one of the people who popularized dubstep, but the song they chose is his latest release "Otto's Chant" which is tech house, not dubstep. Skream hasn't even made any dubstep in quite a while.
Somebody could probably write a program to generate new EDM sub-sub-genres. Somebody probably already has.
Good problem for adversarial learning. Train one ML system to rate EDM, trying to match some metric like total sales. Second system tries to generate EDM which gets high scores from the first system.
One of my favorite things to do is listen to music in foreign languages. (Turkish pop is my favorite. Check out Tarkan or Ismail YK if your ears are curious.) This site is a fantastic way to discover more music from around the world. Great resource!
Lots of other cool Spotify-scraping projects by the author at the bottom.