Autechre production methods speculation

  On 8/28/2016 at 10:09 PM, Zeffolia said:

 

  On 8/28/2016 at 3:10 PM, xox said:

zeffolia, this is very interesting, pls continue.

 

btw...would it be easier to implement AND manipulate all of this in max environment or by using something like supercollider?

 

One other obvious thing I've really been noticing like a lot more recently is subtle frequency envelopes.  Maybe they don't use them and I'm just hearing things but I never really noticed them too much until a while ago.  Everything after Oversteps strategically makes use of starting out as one note and really quickly switching to another to kind of imply a background melody made up of these barely present notes, as well as background drones starting at one and slowly going up to another.  This results in that effect where you can hum a fully fleshed out melody in your head along with the track, but half of the melody you're humming isn't even in the track so you think wtf?  But it actually is there scattered over time so your brain pieces it together and fills in the gaps in future repetitions.

 

Not sure i understand this bit. You got any specific examples?

  On 8/28/2016 at 10:25 PM, misc said:

 

  On 8/28/2016 at 10:09 PM, Zeffolia said:

 

  On 8/28/2016 at 3:10 PM, xox said:

zeffolia, this is very interesting, pls continue.

 

btw...would it be easier to implement AND manipulate all of this in max environment or by using something like supercollider?

 

One other obvious thing I've really been noticing like a lot more recently is subtle frequency envelopes.  Maybe they don't use them and I'm just hearing things but I never really noticed them too much until a while ago.  Everything after Oversteps strategically makes use of starting out as one note and really quickly switching to another to kind of imply a background melody made up of these barely present notes, as well as background drones starting at one and slowly going up to another.  This results in that effect where you can hum a fully fleshed out melody in your head along with the track, but half of the melody you're humming isn't even in the track so you think wtf?  But it actually is there scattered over time so your brain pieces it together and fills in the gaps in future repetitions.

 

 

Not sure i understand this bit. You got any specific examples?

 

The broken apart melody thing is really apparent in draft and oversteps, especially in st epreo and subtle frequency envelopes seem prevalent on O=0.  Having trouble finding concrete examples, sometimes it's easy to be under the illusion of a frequency envelope when it's really just cutoff.  But I know I heard tons when high last month 

  On 8/28/2016 at 2:27 PM, psn said:

Generating motifs that are actually interesting/engaging is something that needs some clever thinking.

Yeah this is an issue

My thinking is that to start out you could create a main melody, then manually create a motif you want as well.  And then create a motif generator which generates that motif in a generic way that can be applied to other melodies as well.  Then use the motif generator on top of a different melody and see if the result is desirable

You could even turn this into a machine learning classification problem (this one's a little advanced)

-Design a generalized motif generator to be controllable in how it generates motifs via N main parameters

-Use a support vector machine (SVM) binary classifier taking vectors of length N.  Let's call this SVM A

Note, binary classifier support vector machines are methods of classifying input vectors of data as being one category or another.  So you can give it a vector of values and assign it a tag of category1 or category2, then after sufficient training it should be able to make the decision for itself of whether a vector is in category1 or category2.  We will use this to train a motif generator

-Take M melodies you made by hand, and for each one

    -Generate N random motif generator parameters

    -Apply the motif generator with those parameters to each of these M melodies

        -For each output

            -Listen to the result, and if it's pleasing to your ears give it a pass (true), and if it's not give it a fail (false)

            -Train SVM A on this vector of N motif generator parameters, with a tag of your true or false designation (category)

Keep doing this over and over on more input melodies and more random generations of parameters, and then presumably this SVM will be trained to categorize motif generator parameters as good or bad, based on your manual training earlier listening to the input

I won't go into more details but basically you can keep doing this over and over in more layers, taking accepted motif generator parameters from your SVM A, and feeding them back into another SVM B but this time being more critical of which motif generators you accept, and you can filter out the worse ones each generation.  This is kind of a genetic algorithm method and it's probably really roundabout (people with more knowledge of ML could easily refine this into a better system probably using some other classifiers) but in the end you will get motif generators that should work on arbitrary input melodies and sound good. 

Careful not to over-train though or they will become generic and bland (this is called over-fitting, which I also referenced earlier in my suggestions to avoid letting your Markov chains have too small of a vocabulary and too high of a degree because then it just starts to copy your input melodies exactly.  Let them be a bit crappy for cool unexpected results.  

Actually sorry that additional layer method is kind of retarded, you can just be really picky with which motif generator parameter vectors you label as true (good sounding) when training the initial SVM and you'd get the same result of good soundingness.

Wow, I'm going to have to block out some time and really go over your posts Zeff. Looks like some great info.

  On 8/27/2016 at 10:39 PM, misc said:

man the visual interface of this is so satisfying to watch

Not as cool, but kind of reminds me of this iPad app called Musyc. It let's you do lots of physics based stuff, but alas with preset sounds. They had said MIDI support was coming but that was awhile ago.

https://www.youtube.com/watch?v=MKhEDEeAP-Q

  On 8/29/2016 at 4:43 AM, weakmassive said:

 

  On 8/27/2016 at 10:39 PM, misc said:

man the visual interface of this is so satisfying to watch

 

Not as cool, but kind of reminds me of this iPad app called Musyc. It let's you do lots of physics based stuff, but alas with preset sounds. They had said MIDI support was coming but that was awhile ago.

https://www.youtube.com/watch?v=MKhEDEeAP-Q

 

Things like this are really cool.  Making the graphics more intense could make it be like real-time generation of abstract music videos to accompany the sounds

I think they're too chaotic to create interesting music though and they don't allow enough control over what's going on

There's lots of sounds that sound like deep clicking noises or like something grinding over a washing rack.  Used all over c7b2

I'm assuming this is just heavily distorted low frequencies as opposed to repeatedly triggered sounds?  Anyone know what I'm talking about?

yes I know exactly what you're sayin..it's all over AE_live and el seq...like a super low frequency sawtooth or triangle...click clicking away...crossing through the rhythm/tone barrier. how they get it to sound so..er.. rich I don't know.

really been enjoying your posts in this thread Zeff

I very humbly think that that isn't a sawtooth but a wavetable oscillator, looping a spikey sample of some sort.

Lots of sounds seem like granular synthesis on a sample of a transient, then compressed so it becomes like a drone of transient grains or a farty noise.  Then you can very easily modulate this sound by moving the location in the sample from which the grains are being selected, forward in time to be less farty and more droney.  A subtle LFO on grain selection location would produce more depth.  Then add in really subtly mixed reverb or resonance.  

Also noticed that really sharp and heavily compressed attacks quickly fading into the normal waveform can give the illusion of really heavy synth slabs of sound.  But this is obvious.  

Then also granular on samples of transients, with extra bursts of release drum-echo-sound and reverb upon the last sharp transient grain before the release of the tone, gives the mental illusion that the preceding sounds were thicker and fuller when you hear it end.

And granular in general makes it extremely easy to modulate stereo width without dumb tricks.  You can literally just make the selection of grains diverge in each channel, and it becomes maximally stereo widened without any weird EQ or phase based tricks.  Can do this really subtly with envelopes on stereo width where a synth stab starts as mono and sharply increases in stereo width, fizzling out into a thick reverbed and echoed tail release.  This seems really common in recent ae_live/elseq unless it's a different method I'm hearing

At the end of the day threads like this are relatively useless

It's like studying guitar academically and expecting to be able to play it the way (insert guitar player here) plays it just because you can speculate on the individual techniques they use (and are probably way off anyway)

Still fun

Yeah possibly granular, possibly also extreme frequency modulation clicking in and out of audible range

Extreame commb filtering with compression gives a similar affect to grain shifting but with a lot control.

I'm steaming right now (in that full on positive way)  and finally have the balls to post in this thread and say Zeffolia your posts here are just brilliant. Probably just because of the arrangement of words. Its almost like science fiction poetry. Not keen on the word 'farty' though. Still 10/10 (even though I haven't got a fucking clue what your on about lol).

Ergh I'm gonna wake up slightly sober in the morning and realise I posted in the Elseq Production Methods Speculation Thread and freak out

  On 8/31/2016 at 7:34 AM, Zeffolia said:

There's lots of sounds that sound like deep clicking noises or like something grinding over a washing rack.  Used all over c7b2

 

I'm assuming this is just heavily distorted low frequencies as opposed to repeatedly triggered sounds?  Anyone know what I'm talking about?

 Maybe just use a very low clicking noise and put many of them manually over a very very fine grid (say 1/128), then you can make sure every click is in time :) laborious but will most probably work

Subtle distortion modulated in intensity by envelope follower to detect transients and make it stateful and alive sounding.  But this is probably the same thing as enveloped distortion anyway.  Though it would be more automated.

Gate to distortion sidechain?

That sounds like a better way to describe it yes

I swear I heard trailing echoes with a pre-delay on them which extended before that initial sound which generated the echoes themselves, in Exai somewhere last night while I went to sleep listening to it

id like to read more of this pls

How the hell did they make all the weird noises in the end of squeller?

is it just some weird digital delay thing?

It could be boss rdd-10 + boss rsd-10 or it could be automated quadraverb

and so on