Using Dithering In GarageBand ???

[musicworld]

Hi

How does one apply dithering in GarageBand ? where do you go to enable the feature ? or does it even exist in GB.

[tony tony chopper]

I don't understand why for so many years people are still hinted that they need to care about dithering, but this is something you should watch
http://www.youtube.com/watch?v=cIQ9IXSUzuM

[V0RT3X]

I don't understand why for so many years people are still hinted that they need to care about dithering, but this is something you should watch
http://www.youtube.com/watch?v=cIQ9IXSUzuM

Wow this is impressive!

Also.. I want those oscilloscopes..

[schnapsglas]

I have some problems with methods used here, but I don't have the time to comment. Hopefully, someone else will do it

[earlevel]

Hi

How does one apply dithering in GarageBand ? where do you go to enable the feature ? or does it even exist in GB.

Could you elaborate on what you want to do?

In general, you should dither when saving to 16-bit files. That usually means your final mix, exporting to 16 bit. If GB doesn't dither to 16 bit (yikes), then you could just open a 32-bit (float) or 24-bit mix in another program that will dither to 16 bit.

If you're talking about dithering to 24-bit, that's a practice of questionable value. I'll probably have an article for my website sometime soon on the subject, as I've had lengthy private discussions with a mastering engineer on the topic recently, and would like to put up some math and examples.

[schnapsglas]

Hi

How does one apply dithering in GarageBand ? where do you go to enable the feature ? or does it even exist in GB.

Could you elaborate on what you want to do?

In general, you should dither when saving to 16-bit files. That usually means your final mix, exporting to 16 bit. If GB doesn't dither to 16 bit (yikes), then you could just open a 32-bit (float) or 24-bit mix in another program that will dither to 16 bit.

If you're talking about dithering to 24-bit, that's a practice of questionable value. I'll probably have an article for my website sometime soon on the subject, as I've had lengthy private discussions with a mastering engineer on the topic recently, and would like to put up some math and examples.

Dithering to 24, you simply should not do, in my opinion. Unless it makes you feel better... and sure that that is the last conversion it will go through.

[earlevel]

Dithering to 24, you simply should not do, in my opinion. Unless it makes you feel better... and sure that that is the last conversion it will go through.

Agreed, if it wasn't clear from my comment that the practice was "questionable". Still, the idea really bugs some people...most of which probably don't realize that most Pro Tools (HD/TDM) plug-ins don't dither their 24-bit output...

[tony tony chopper]

if you believe you need to dither to 16bit, post that one bit of audio that will convince everyone that it makes an audible difference. It has never happened & never will, because it simply doesn't.
(but you can always dither, because it's not harmful either)

All of the websites that prone (or sell) dithering tools have always done the same thing: they put less than 12bit worth of audio in an audio file, & tell you to crank up your volume until you can hear it. Effectively proving that you need dithering.. to 12bit.
So if your song ends with a nice reverb tail, and you believe that listeners are gonna raise their volume a lot to keep hearing that reverb tail when it goes inaudible, you need dithering.


I see earlevel above is honnest, on his website he uses 7bit audio to demo dithering.
Dithering (& ADPCM) was very important for handheld consoles that were dealing with 8bit audio. But 8bit audio is dead, not used in anything anymore (to my knowledge). I've heard some really impressive demos of ditherers btw, but ironically they were made while 8bit audio was already dead.

Same deal with graphics, with the difference that most graphics SDKs work with 8bit RGB, and 8bit per (non-linear, in this case) component is just not enough for our perception, we'd need 9 or 10bits, so things like gradients really need dithering to avoid banding. With 16bit graphics, same deal, you wouldn't need to bother with dithering at all (only when converting to 8bit).

[earlevel]

if you believe you need to dither to 16bit, post that one bit of audio that will convince everyone that it makes an audible difference. It has never happened & never will, because it simply doesn't.
(but you can always dither, because it's not harmful either)

Thanks, Tony, for being so bold as to post the "emperor's new clothes" point of view. I don't disagree with your point that, broadly, it's not going to make a difference. Not in a world that listens to mp3s at least.

But the bottom line is that you can hear it easily, under the right (read: "largely unnatural") conditions, particularly with manufactured signals. So it's a must-have if you want to sell to recording and mastering engineers etc. These guys do NOT want to pay for a lot of expensive equipment and hear such an easily-avoiable flaw—it doesn't matter that, in real situations, no one would hear it. And there's no real cost if you're doing it when you are exporting a 16-bit audio file. (I'm assuming everyone's working in 24-bit or greater till that point anyway.) So, the bottom line is that everyone should dither to 16-bits if it's pro-level stuff, because it's silly not to. Again, any pro-level stuff is going to be working at a greater bit-depth than 16 for several reasons, and we're usually talking about saving a 16-bit file at this point.

So, I consider the question of "necessity" in dithering to 16-bit a moot point. You can argue yes it's important, or no it's not important—in the end are you going to make a high-quality digital audio thingie that can put out 16-bit audio and you intend to sell to recording studios a not do it? Risk having a review in Sound on Sound with pumps a tiny sine wave in and gets something that sounds like a square wave out?

So no, I'm not going to go to the effort of making a test case, because it's easy enough to dismiss. But I can give people an idea of what "perfect" LSB-level audio sounds like. I generated this as part of a lengthy private discussion I had recently with a master engineer who feels that even 24-bit audio should be dithered, at ever source of truncation (I think he was horrified when I told him I'm pretty sure that few Pro Tools plug-ins, which route 24-bits, don't dither. The ones I wrote certainly don't).

OK, what I mean by "perfect": This file is basically a mono 24-bit WAV file, a square wave starting at Nyquist and dropping for a ways (I didn't pay attention how far—one constraint was keeping it a comfortable size for email). So it starts off with 16th bit set for one sample, then negated for one sample—repeat a while...then two up, two down for fewer times, 3 by 3, 4x4...so, the frequency resolution is poor at the start and gets better as the frequency drops (this is a "divide by n" oscillator—that's what they do), but it has the advantage of having no quantization error or oscillator jitter.

The purpose is to get an idea of how loud the signals are that we'd be dealing with, in comparison to music. Using your studio speakers or headphones, crank up some music to a loud but typical monitoring level. Then shut that off and listen to this file. You will have no trouble hearing it, and there is no mistaking the square-wave quality.

Now, consider that a sine wave of around the same level will end up sounding something like that square wave. But if you dither, it will sound like a sine wave and a barely-detectable amount of noise. If you use noise-shaped dither, it should sound like just the very quiet sine wave. Sure, what fool would allow such a tiny signal, sweeping around to draw attention to itself, in a momentarily dead-silent portion of the music or fade-in or fade-out. Normally you won't hear this. But again, why not.

16th bit

I'll give you a 24-bit version of the same thing. Good freaking luck hearing it without a lot of amplification, plus cryogenics to kill the thermal noise:

24th bit

In case you want to hear what you're listening for better, here's a version for the 5th bit. All of these are 24-bit files, the only change is the amplitude (you might want to back off the volume if you cranked it for the previous waves):

5th bit

[tony tony chopper]

it doesn't matter that, in real situations, no one would hear it. And there's no real cost if you're doing it when you are exporting a 16-bit audio file.

Although I really doubt that a studio would even notice dithering or lack of it (& why would they even want to be provided 16bit audio to start with?), that's where we agree, dithering won't hurt anyway.
..except that it kinda hurts, because a lot of people out there are wondering why their sound sucks (generally cases of muddy mixes), and as they hear about dithering everywhere, that's what they assume they need to get a better sound, dithering. Sure, they only wasted their time, except if they start looking for better ditherers because they didn't notice any difference

There is so much to care for in audio, problems that are really audible & matter a billion times more than dithering..
(the most ironic is that all "real" audio recordings have a noise ratio so so much higher than this)

Btw, my reason not to bother with dithering to 16bit is your reason not to bother with dithering to 24bit. We just don't agree on the quality of human perception, it's not really about dithering. I facepalm whenever I see an ugly banded gradient & people don't care (this forum is full of banded gradients btw, deep enough to make the banding not so annoying but still..). But I've never analized the RGB values of a 16bit image to check if there's some banding I can't see. And if I have to post-edit the image, I want the original precision anyway, not a 8bit version, dithered or not.

[schnapsglas]

I don't understand why for so many years people are still hinted that they need to care about dithering, but this is something you should watch
http://www.youtube.com/watch?v=cIQ9IXSUzuM

OK, as I hinted, I will voice my objections.

What he says is true. Lot of people seem to think digital samples are discontinuous, when they are in fact continuous. That is definition of a sequence.

However, he shouldn't do that test with a sine wave. It is simple mathematics that if you get closer to Nyquist (which is often stated incorrectly) you can't get even a "reasonable" bijection from the waveform to the digital samples. You will get A/D -> D/A pretty clean with sine wave, but it will not be case for other waveforms. (which is THE POINT of Nyquist-Shannon!)*

Yeah, he is right, but for god's sake, let's do this properly.

*If I'm wrong, please correct me.

[tony tony chopper]

*If I'm wrong, please correct me.

I don't see how it's right

At least in theory. In practice you need quite a lot of samples (thus a big latency) to interpolate a sinewave that's near nyquist, but in theory you obviously can.

[schnapsglas]

*If I'm wrong, please correct me.

I don't see how it's right

At least in theory. In practice you need quite a lot of samples (thus a big latency) to interpolate a sinewave that's near nyquist, but in theory you obviously can.

What I am saying is getting a sine wave when frequency is strictly less than the Nyquist frequency with no harmonics is a trivial point since you are just reconstructing the fundamental. It is rather a vacuous thing to say.

The test should be done with say, a square wave at 2^2*3*5^2*7^2/2 = 7350hz. If it gives something very close to sine wave at 7350hz, now that would be something worth looking at.*

*Again, please correct me if I am wrong. Nowhere near my area.

[earlevel]

The test should be done with say, a square wave at 2^2*3*5^2*7^2/2 = 7350hz. If it gives something very close to sine wave at 7350hz, now that would be something worth looking at.*

OK, I assume you're talking about 44.1 kHz sample rate...and it seems that you picked 7350 Hz so that the square wave would have first harmonic at 7350, third at Nyquist, and odd harmonics continuing above Nyquist. So, if you're recoding (sampling) this, the anti-alias filter ditches everything except the fundamental (within the specs of the converter, but the alias components should be insignificant). So, you play back a sine wave at 7350.

Or, if you mean a calculated square wave, that would just be the 7350 sine wave.

Am I following you correctly?

(BTW, at some point it's possible that a staircase is getting output, or something like it, but it gets smoothed by the reconstruction filter. Conceptually, samples are impulses, but it's easier to sample and hold on the output. It makes the frequency response wrong (the highs roll off), by that can get fixed in the reconstruction with a little high-end boost.)

[schnapsglas]

The test should be done with say, a square wave at 2^2*3*5^2*7^2/2 = 7350hz. If it gives something very close to sine wave at 7350hz, now that would be something worth looking at.*

OK, I assume you're talking about 44.1 kHz sample rate...and it seems that you picked 7350 Hz so that the square wave would have first harmonic at 7350, third at Nyquist, and odd harmonics continuing above Nyquist. So, if you're recoding (sampling) this, the anti-alias filter ditches everything except the fundamental (within the specs of the converter, but the alias components should be insignificant). So, you play back a sine wave at 7350.

Or, if you mean a calculated square wave, that would just be the 7350 sine wave.

Am I following you correctly?

(BTW, at some point it's possible that a staircase is getting output, or something like it, but it gets smoothed by the reconstruction filter. Conceptually, samples are impulses, but it's easier to sample and hold on the output. It makes the frequency response wrong (the highs roll off), by that can get fixed in the reconstruction with a little high-end boost.)

Yes, you are. I mean as in the video, using analog signal to generate square wave at 7350, and then the second oscilloscope after being converted digital and back should show 7350 sine wave, hopefully. I thought that would be the way to show A/D D/A chain shown in the video isn't faulty and makes better case.

I hope I am not saying anything stupid, because all I know is how theoretically a wave is reconstructed from samples.