TL;DW: Deterministic dither uses Benford Realness calculations for each sample.

NaturalizeDither

Last and definitely not least in the dither-fest: Naturalize! This dither is often considered best by listeners. It works on every genre, at every sample rate. It adds no tonal color to the audio, and the noise it produces is quiet… but more than quiet, the noise of Naturalize has a fugitive quality where it seems to ‘hide’ behind the reverb tails, and once you start listening to the music it’s difficult to hear there’s any noise present. The sense of reality produced by the music is compelling enough that it doesn’t draw attention to the noise floor at all.

How can this be? I kept the secret for eight years but now that Airwindows is supported by a Patreon, I’m free to communicate how I did it, because my compensation is basically ‘the number of people out there who think I ought to keep working in this field’, and the more of those the better. I don’t have to keep Naturalize just to myself anymore, because it helps the Patreon when I’m generous. So, go ahead and download and use it for free (it’s running at 24 bit, use BitShiftGain to get 16 bit out of it) and here’s how Naturalize works.

Benford’s Law is a statistical technique: basically, it says that in lists of natural data, if you get rid of all the leading zeroes, the first digit other than zero is most likely to be ‘one’. That’s no matter what scale the number is at: no matter how many zeroes you have to take away to get to the first number, it’s more likely to be ‘one’ than, say, ‘nine’.

In fact, it’s ‘one’ thirty percent of the time, and the other leading digits are progressively rarer. How much rarer? I’ll give you a piece of the Naturalize code and the numbers I’m using.

byn[0] = 1000;
byn[1] = 301;
byn[2] = 176;
byn[3] = 125;
byn[4] = 97;
byn[5] = 79;
byn[6] = 67;
byn[7] = 58;
byn[8] = 51;
byn[9] = 46;

This works on sample data that varies in logarithmic scale: fortunately, audio is already that way. Something like a reverb tail is perfect for Benford realness calculations! As the sound dies away, it SHOULD fall into just this sort of pattern: millions of sample numbers, scaling down to tiny near-zero values. It ought to be the perfect picture of Benford’s law.

Here’s what Naturalize is. Every sample can be truncated in two directions: up or down, to the nearest value. So we do both, and then we run the Benford’s Law calculation on each, and we see which choice will get us closer to that perfect ‘realness’ distribution.

Then we just choose that direction :) that’s it! Each sample, we always choose the direction that will bring us closer to Benford’s Law.

This works on a statistical basis, to produce a collection of sample values that hold to what you’d get if you directly sampled ‘reality’. It erodes unnatural qualities in the digital audio, and it’s pleasing to listen to. There’s no excess energy to worry about: it’s always within one least significant bit of the target audio. And I might be able to improve upon it (I initially remembered it as having no noise sources at all, but it’s using some noise to shake up the values a bit for better distributing) but as 2017 begins, doing this gives you the best sounding dither in the world. I’d be happy to see if I can make it even better: support my Patreon, after all I came up with this thing in the first place, stands to reason I can develop it even further. Maybe I’ll make it entirely deterministic, with no noise at all. I’m sure I tried that already, but sometimes persistence pays off.

Happy holidays, I’ll have a little present on Sunday for you, and I hope things are well :)

TL;DW: The testbed for the TPDF-based dithers. Flanger dither!

NodeDither

And just one more dither plugin before the eagerly anticipated Naturalize… (because who will care about these then?) …NodeDither, the testbed upon which PaulDither and TapeDither were developed!

I still think it’s good to have those as separate plugins (for simplicity and purity of concept) but since the concept is ‘TPDF made up of a sample of noise, and then a delayed version either normal or inverted’, how can I not give out the tool I used to find things like TapeDither? This was coded to do something that wasn’t possible: put a big delay in so there’s a ‘cancellation node’ at around 1.5K, making the noise quieter there.

Trouble was, all you get is comb filtering and no nice clear ‘quieter’ zone. That’s why I experimented and found that four samples delay (inverted) gives you TapeDither, a nice softer noise. One sample inverted gives you PaulDither (in other words, a very bright one-pole highpass). Two samples gives you a neat silky texture which is still very bright. Eight samples gives you a thing kinda like TapeDither but even darker and beginning to sound flangey… you might call it CassetteDither, or maybe ‘TapeDither for 96K’.

Call it what you like, because NodeDither lets you use all those and anything else, up to really extreme delays that give heavy, heavy comb filtering (flangeyness). You can reinforce lows rather than cancelling them through out-of-phase, and dial in dark moody dither noises that fit better with your project. You can tune the flangey note of the dither to your track. All manner of silly things can be yours with NodeDither :)

And yes, you can make your noisefloor a flanger. I’ve demonstrated in the video how to do this with automation. If people really REALLY need a flange dither, I can make one later, but Naturalize will not be delayed any longer and is due to come out this Wednesday. But hopefully the possibilities of NodeDither are worth having! Or, put it this way: if NodeDither is for you, you know who you are and you’ve probably already downloaded it and are playing with the controls and thinking of how it would affect your creations.

Because, bottom line: this is still technically correct TPDF dither. It’s just also an instrument, now. Have fun playing it :)

The concept of a standard TPDF dither that’s also an instrument was brought to you by my Patreon: if enough people get involved in this, I can live or even invest in further experiments. Airwindows Console4 only happened because I was able to study a real analog desk (an Allen & Heath) and drive it with a nice multichannel DAC (a MOTU 16A) and study the results. The more money I make, the grander the experiments can become… and of course there’s extra stuff on my YouTube channel, live streams that are only announced to my Patreon patrons. (I suppose you could just hover over the channel, or search it, but I’ll make a point of alerting the Patreon crowd when I do something like that.)

The most recent (okay, first) live stream was hacking a Xiaomi Yi action cam, and making the mic better, so there was some audio content! It also involved un-gluing the lens so it could be removed, and installing a deadcat wind sock for outdoor use. I’ll try to keep a steady flow of unusual things happening on the Airwindows youtube, though most of them cannot be announced in plugin forums. Next, hotrodding my already hotrodded Squier J-bass to add a Artec mudbucker pickup! This was inspired by Rick Laird of Mahavishnu Orchestra. Only then did I learn this was also the pickup complement favored by Cliff Burton… I’ll post to patrons when that livestream is ready, and you can see the radical things I’ve already done to the J, and hang out while I do the work :)