The No.1 Website for Pro Audio
 Search This Thread  Search This Forum  Search Reviews  Search Gear Database  Search Gear for sale  Search Gearslutz Go Advanced
Good dither practices, what are yours? Dynamics Plugins
Old 10th March 2017
  #391
I pressed "export" in wavelab. I'm sure you can do that, too. Come on...
Old 10th March 2017
  #392
Looking at / listening to the error signal is a smart thing to do, although not many toolsets support it conveniently. I agree that if you see a single spike in the autocorrelation function, then your quantization noise is spectrally white. The other thing you should do is look at the error amplitude histogram. You'll find the peak errors to be higher than in the TPDF case. Maybe not audible, but objectively sub-optimal. The main thing to check here is that the error mean is zero, though what you should really verify is that the output mean is zero. (DC offsets cause clicks and pops downstream.)

Still, assuring spectrally white error doesn't mean that chrism can't hear it, because the noise he used as an input signal might have been colored, as is most real-world noise.
Old 10th March 2017
  #393
Gear Maniac
 
Yuri Korzunov's Avatar
 

Quote:
Originally Posted by David Rick View Post
That's what it takes to make the statistical moments of the quantization error uncorrelated with the input signal. No other noise will do that perfectly, but non-ideal noise can lower the input->error correlation and higher amplitude noise will usually lower the correlation.
If remove musical signal, noise stay as is.

Noise generated by independent (on music) source.

So what is correlation may be there?

For correlation music and dithering noise need apply special action.

For decorrelation noise and music need do nothing.
Old 10th March 2017
  #394
Quote:
Originally Posted by Yuri Korzunov View Post
Noise generated by independent (on music) source.
So what is correlation may be there?
For correlation music and dithering noise need apply special action.
For decorrelation noise and music need do nothing.
Not true. The inputs (music & dither) are independent but the error at the quantizer output is correlated with the music. That's the fundamental problem! But by choosing the correct kind of dither (TPDF of precise amplitude) we can force zero correlation in the first two moments. Thus the mean error and its power spectrum become independent of the music. (Those two moments seem to be all we can hear.)

Read this:

Quantization and dither: A theoretical survey

David
Old 10th March 2017
  #395
qwe
Lives for gear
 

Quote:
Originally Posted by David Rick View Post
Not true. The inputs (music & dither) are independent but the error at the quantizer output is correlated with the music. That's the fundamental problem! But by choosing the correct kind of dither (TPDF of precise amplitude) we can force zero correlation in the first two moments. Thus the mean error and its power spectrum become independent of the music. (Those two moments seem to be all we can hear.)

Read this:

Quantization and dither: A theoretical survey

David
Or see https://www.gearslutz.com/board/12494897-post378.html for a link to a free copy of the paper...

Question is, what's the third moment and why doesn't it (psychoacoustically) matter?

Furthermore, IIRC it does matter once we start processing the signal in certain ways?
Old 10th March 2017
  #396
qwe
Lives for gear
 

Quote:
Originally Posted by FabienTDR View Post
Sure, I'll be happy post some plots later, discussing results one by one. But I'd really prefer if ppl would try this out on their own. It's more neutral

For now, you just have to load up a FFT analyzer, set the window size as high as possible and use strong windowing. Set the floor value very low, -180 or so.

Then, simply scan through the files. Start with the original. It contains 5 segments, let's try to preserve them:

1. Stereo dirac (irrelevant to this discussion)
2. A sine pair
3. Seven sines
4. A sine pair, with natural noise floor.
5. Seven sines, with natural noise floor.

The other two files suffixed "truncatedTo16" or "truncatedTo24" have simply been exported, without any explicit dithering.
OK, I just tried the "truncatedTo16," I can see what you're talking about comparing 2/4.

Are you sure 3/5 aren't the same?

It'd be better to upload your 'brown noise,' or specify how it was generated, then we could knock up our own tests. :-)
Attached Thumbnails
Good dither practices, what are yours?-dithertestsignal_v1_truncatedto16_spectralplot.jpg  
Old 10th March 2017
  #397
precise amplitude? AFAIK, these papers all talk about minimum amplitude. That's a difference. It would make no sense anyway, as this would imply that not all signals can be dithered, but only those with an exact amount of noise absence, we still want to add some, right? (imagine a signal that already contains noise, so your levels are potentially off). Is that what you are saying? It makes no sense to me.

Again, it's all about making the error fully random (i.e. lifting the noise floor high enough), then, no partials can appear, by design. Period. That's rather simple and proven mathematical relation. Random signals are still random after most forms of processing. Noise cannot have harmonics, impossible. That's dithering. You make the error random, so that no partials appear. Coloured noise is sufficient to prevent the generation of partials (simple fact: noise shaping is nothing else than noise colour). No need for further complication. As many ppl stated here: too much dithering noise doesn't hurt either, and typical signal noise floors are MUCH higher anyway.

In doubt, take any wideband noise you like, blue, pink, brown. And find the most radical distortion you can find. Then try to distort it audibly or measurably.

All these dithering papers have a strong focus on the reduction of THD (i.e. making them THD optimal). Keep this in mind. These are minimum values when you want to truncate a sine (the worst case signal for truncation, but very unlikely with music) and keep THD at optimum. Did you see my Seven Sine example without any noise, and how it behaves when truncated? Can you detect any quantization distortion? Isn't that, let's say, slightly "suspicious"?

Last edited by FabienTDR; 11th March 2017 at 12:12 AM..
Old 10th March 2017
  #398
Just a moment!

Quote:
Originally Posted by qwe View Post
Question is, what's the third moment and why doesn't it (psychoacoustically) matter?

Furthermore, IIRC it does matter once we start processing the signal in certain ways?
First moment is the "expected value" of the error. We can write this as E{ e(t) }. We want this to be zero at all times, t.

Second moment is an expected value like this:
E{ e(t1) * e(t2) }.
When t1=t2, it's the instantaneous error power. We can't make this zero, but we want it to be independent of the input signal, otherwise we'll hear the noise floor get modulated by the music. When t1 and t2 are different, this is called an "autocorrelation". Normally, this is "stationary" meaning it only depends on the time difference, (t2-t1). Take its Fourier transform, and you've got the noise power spectrum. In the simple case, this is white (flat), but in a noise-shaping quantizer it can have some other frequency response. Wannamaker showed how to make it "minimally audible" according to threshold of hearing response.

Obviously, the third moment has this mathematical form:
E{ e(t1) * e(t2) * e(t3) }
What's it mean in practice? It's hard to say, because apparently nobody can hear it. But if it worries you, you can drive it to zero by using third-order dither, which is a sum of three independent RPDF noises. (Hey Chrisj: Try it and see if you hear a worthwhile difference! But I predict you'll only hear a slight increase in the noise floor.)

Few statisticians bother with moments higher than two, because two moments completely describe anything with Gaussian statistics, and Gaussian is the most popular PDF in the universe due to the Central Limit Theorem.

In fact, the more I think about it, the more I think we shouldn't do third-order statistics at all. We apparently can't perceive them, and if there exist extraterrestrials who can, they probably think third moments are some kind of religious abomination that would justify them destroying our planet. Why do I think that? Well, second order statistics have a connection to the work of Joseph Fourier (see above), and our new EPA administrator is so determined to deny the work of Fourier that he's already willing to risk destroying the planet!

I shall not elaborate on that previous statement for fear of being banned from this forum by introducing politics. Google Joseph Fourier and figure it out for yourself.

David L. Rick
Seventh String Recording

Last edited by David Rick; 10th March 2017 at 11:41 PM..
Old 10th March 2017
  #399
Quote:
Originally Posted by qwe View Post
Are you sure 3/5 aren't the same?
Damn, you are right. Sorry. Indeed, noise stops at some point during 4)

I fixed all files.

For brown noise, just integrate white noise. It's just a crude assumption for our case, most real world signals and circuits will show a mix of white and brown.
Old 11th March 2017
  #400
Quote:
Originally Posted by FabienTDR View Post
precise amplitude? AFAIK, they all talk about minimum amplitude. That's a difference. It would make no sense anyway, as this would imply that not all signals can be dithered, but only those with an exact amount of noise absence, we still want to add some, right? (imagine a signal that already contains noise, so your levels are potentially off). Is that what you are saying? It makes no sense to me.
Fabien, it's the dither amplitude that should be precise -- depends on the size of the quantization steps. Here's a graph for RPDF dither; for TPDF dither it looks very similar, but wider notches starting at 2 LSB's rather than at 1.

David
Attached Thumbnails
Good dither practices, what are yours?-peak-error-curve-rpdf-dither-varying-amplitudes.jpg  
Old 11th March 2017
  #401
I understand this part. It is the minimum level of randomness needed to allow the mechanism to function (for a sine, not in general).

Again, otherwise, it would be impossible to successfully dither noisy signals. That's straight ridiculous. Overcomplicated, bureaucratic thinking about something trivial.

Dithering is important. But with music signals, it's already given in most cases. It's called a noise floor. There is no difference, noise is noise is noise. Otherwise, "this toxic, unditherable noise" would clearly show partials when truncated. It can't happen. Don't you see the logical problem with your interpretation?

Are you seriously saying that dithering will fail if noise is too strong?!

Dithering is no rocket science. Do not overcomplicate the matter posting all sorts of basic dithering papers. Show partials popping up somewhere, if you don't see them (or prove their existence in a way or another), they most likely simply belong to the incredibly powerful subjective imaginary power of music.

We are moving below the threshold of audibility, and far above the minimum level of dither. Seriously, why this incredible run trying to enforce some sense into dithering, while most of my basic theoretical and practical examples clearly point to the opposite (which, btw, nobody seriously questioned so far).

Beyond all the technical arguments and examples I posted, I'm still not aware of any serious study being able to even show human perception of dithering or the absence of it at 16bit or beyond. What would you do the day you find out that the relevance of dithering for normal music programs and formats was just imagined? Technically irrelevant for modern (16bit+) music production. Change your mind? lol

Insisting on adding -140dB of noise to a -60dB noise floor for dithering is ridiculous (that's decibel!). Draw the data linearly on a looooong paper-sheet to get a feel of the scales we are talking about. Microbes vs elephants.

IMHO dithering in modern audio production contexts is a dogma (the ferocity and ignorance toward scale and simple logic in this debate, and the habit to make it a personal crusade fits well into it). Not more, not less. It's neither a qualitative factor nor a blocker. It's an irrelevant, academic "problem". ...until someone proves otherwise... (but I don't see this)

Last edited by FabienTDR; 11th March 2017 at 01:24 AM..
Old 11th March 2017
  #402
Quote:
Originally Posted by FabienTDR View Post
I understand this part. But it is the minimum level.

Again, otherwise, it would be impossible to successfully dither noisy signals.
Still some disconnect here. Proper dither level does not depend on the noise amplitude or on the signal amplitude in general. It only depends on the step size of the quantizer. Once you've established that step size, you can double, triple, or quadruple it if you want, but don't go multiplying 1.5 unless you're doing is as some kind of an effect.

Quote:
BTW, you further find several commercial dither products offering various forms of noise. They balance the distortion vs THD. And ppl like it!
I'm guess you meant to write "distortion vs noise level" here. Certainly one can do that. My DAW permits non-integer dither amplitudes, but I'm guessing more people get in trouble with that feature than use it beneficially.

Distortion is simply signal-correlated error, and you're correct that academic papers always assume it's bad. As audio engineers, we know that sometimes distortion is artistically good. So it would be no surprise to me if certain people are using "improper" dither as an artistic effect.

When you write "More complex signals need less dithering noise, or even none.", I'll walk a little way with you, then stop. It comes down to masking. Some distortion components can be masked by the original signal or covered up by the noise floor. It's very case-dependent, though, and not something you want to count on when defining your standard workflow.

Quote:
Dithering is important. But with music signals, it's already given in most cases. It's called a noise floor. There is no difference, noise is noise is noise. Otherwise, "this toxic, unditherable noise" would clearly show partials when truncated. It can't happen.
Did someone say that? I must have missed it. I did understand someone here to say that he could hear the difference when truncating vs. dithering an input signal which he called "noise". Well, there are many different kinds of "noise", despite your comment above. Often the acoustic noise in concert halls has some tonal components. Electronics have Johnson noise (which I think is white), but also 1/f "flicker" noise (which clearly isn't). Microphone noise is partly due to Brownian motion, but it gets shaped by mechanical resonances in the capsule. We choose to call 50/60 Hz power line leakage "noise" as well. But with the agenda of establishing a safe default for audio processing, it makes sense to treat all these various "noises" as just another signal, and dither them like we do everything else, rather than hoping they'll bail us out when we do something foolish.

David
Old 11th March 2017
  #403
Gear Maniac
 
Yuri Korzunov's Avatar
 

Quote:
Originally Posted by David Rick View Post
Not true. The inputs (music & dither) are independent but the error at the quantizer output is correlated with the music. That's the fundamental problem! But by choosing the correct kind of dither (TPDF of precise amplitude) we can force zero correlation in the first two moments. Thus the mean error and its power spectrum become independent of the music. (Those two moments seem to be all we can hear.)
Exactly. Aim of dithering is quantization noise (a.k.a. "error of quantization") that independent on musical signal.

If dither applied properly, there is misic signal or it absent, it is no matter.

If truncated (rounded) signal correlated with error (quantization/rounding) we can see harmonics of the signal at spectrum.

There is signal - there is the harmonics at spectrum. No signal - no harmonics.

These harmonics are sign of the correlation. These harmonics are part of quantization noise.

If dither applied properly, spectrum of quantization noise is spectrum of white noise, independently amplitude of signal.

Quote:
Originally Posted by FabienTDR View Post
Dithering is no rocket science.
Fully agree.
Old 11th March 2017
  #404
Airwindows
 
chrisj's Avatar
Quote:
Originally Posted by David Rick View Post
I did understand someone here to say that he could hear the difference when truncating vs. dithering an input signal which he called "noise". Well, there are many different kinds of "noise", despite your comment above. Often the acoustic noise in concert halls has some tonal components. Electronics have Johnson noise (which I think is white), but also 1/f "flicker" noise (which clearly isn't). Microphone noise is partly due to Brownian motion, but it gets shaped by mechanical resonances in the capsule. We choose to call 50/60 Hz power line leakage "noise" as well. But with the agenda of establishing a safe default for audio processing, it makes sense to treat all these various "noises" as just another signal, and dither them like we do everything else, rather than hoping they'll bail us out when we do something foolish.

David
THANK you. Also bear in mind we're being 'Gish galloped', and it's exhausting and annoying, and 'loudest most determined shouting' is not actually science. And of course you're right, for some very specific reasons.

The noise I took double blind ABX tests on, and distinguished infallibly from low-bit noise if and ONLY if it wasn't TPDF dithered, was white noise. Chosen specifically because it's an optimal case for 'self-dither' and real-world noise will be substantially worse.

Here's why: if you try to replace the white noise with say brown noise, while it sounds nicer due to all the bass content, it's useless to dither with: it's so lacking in high frequency variation relative to the lows, that the output will effectively get truncation artifacts just because waveforms getting truncated won't switch between adjacent truncation values fast enough to average out the results on playback. There's a reason high performance dithers and wordlength reducers focus so intensely on near-Nyquist frequencies: those are the ones that most effectively approximate output positions between the quantization levels. And that's how this works, even for simple TPDF dither. (and no, you can't dither low frequencies better with low frequency noise because that's not how dither works)

But then some people behave like the noise floor is made of concrete, and impervious to the ear apart from any harmonic partials that can pop up on a spectrum analyzer…

Quote:
Damn, you are right. Sorry. Indeed, noise stops at some point during 4)
I fixed all files.
Fabien, please stop unless you're prepared to be rigorous. All you're doing is killing the thread. Perhaps that's the intention, but it wasn't yours to kill. If you're casually putting up irrelevant (dirac) and WRONG files and still behaving as if you're entitled to browbeat everyone into qualifying themselves by them, when they are broken and need to be 'fixed', we can't proceed on that basis.
Old 11th March 2017
  #405
Lives for gear
 
stinkyfingers's Avatar
 

Quote:
Originally Posted by FabienTDR View Post
In doubt, take any wideband noise you like, blue, pink, brown. And find the most radical distortion you can find. Then try to distort it audibly or measurably.
for your amusement i attached a file of gaussian noise vs. distorted gaussian noise.
i was able to hear a difference, i got 100% on all ABX attempts. (it was pretty easy, actually)
sample peak and RMS values match closer than 0.0001 %, LUFS, TP, and all those other trivial measurement are within a few tenths of a decibel. so there is a measurable difference as well. and it is not just because of the added DC...i have tested with that removed and then level matched again and the results are the same.
i did do this quickly so there is a possibility i made a mistake somewhere, so i encourage you to try for yourself.


our brains do not perceive "noise" as "random". the brain tries to make some sense of it. the same way we "see things" in an ink blob, we "hear things" in "noise". if you listen to the same noise long enough, you will find/make recognizable/repeatable patterns. we even synthesize our own "tones" out of the "noise". once you have established a firm reference point, it's easy to hear even subtle changes if you are actually listening and not just hearing.
so in a sense this does have to do with a "magical perception" and hearing things that aren't "really there".
the sentiment of "noise is noise" will only determine the depth of your own perception, not others'.
sorry for the crude attempt at an explanation of the science...hopefully i don't feel as dumb as i look.

I no naka no kawazu taikai wo shirazu...
Attached Files
File Type: zip noise.zip (93.9 KB, 4 views)

Last edited by stinkyfingers; 11th March 2017 at 04:17 PM..
Old 11th March 2017
  #406
Lives for gear
 
Arksun's Avatar
Quote:
Originally Posted by Plush View Post
Bickering over a minus 160 dB noise?

Perhaps it will always continue.

Noises out of Germany in particular.

A fetish for exactness.
Exactly, because distortion occurring below the noisefloor of the converter or the threshold of human hearing is very important
Old 11th March 2017
  #407
Quote:
Originally Posted by stinkyfingers View Post
for your amusement i attached a file of gaussian noise vs. distorted gaussian noise.
i was able to hear a difference, i got 100% on all ABX attempts. (it was pretty easy, actually)
sample peak and RMS values match closer than 0.0001 %, LUFS, TP, and all those other trivial measurement are within a few tenths of a decibel. so there is a measurable difference as well. and it is not just because of the added DC...i have tested with that removed and then level matched again and the results are the same.
i did do this quickly so there is a possibility i made a mistake somewhere, so i encourage you to try for yourself.


our brains do not perceive "noise" as "random". the brain tries to make some sense of it. the same way we "see things" in an ink blob, we "hear things" in "noise". if you listen to the same noise long enough, you will find/make recognizable/repeatable patterns. we even synthesize our own "tones" out of the "noise". once you have established a firm reference point, it's easy to hear even subtle changes if you are actually listening and not just hearing.
so in a sense this does have to do with a "magical perception" and hearing things that aren't "really there".
the sentiment of "noise is noise" will only determine the depth of your own perception, not others'.
sorry for the crude attempt at an explanation of the science...hopefully i don't feel as dumb as i look.

I no naka no kawazu taikai wo shirazu...

The only problem is that your distortion adds an enormous DC offset. And that's what you hear and see.

Last edited by FabienTDR; 11th March 2017 at 05:05 PM..
Old 11th March 2017
  #408
My dither practice....Make music and don't worry about dither, because it doesn't matter.
Old 11th March 2017
  #409
Lives for gear
 
stinkyfingers's Avatar
 

Quote:
Originally Posted by FabienTDR View Post
The only problem is that your distortion ads an enormous DC offset. And that's what you hear and see.
I took that into account and I don't believe that's what I am hearing.
What happens when you remove the DC ?
Do the files null ? Are they the same other than the D.C. ?
Old 11th March 2017
  #410
Quote:
Originally Posted by chrisj View Post
Fabien, please stop unless you're prepared to be rigorous. All you're doing is killing the thread. Perhaps that's the intention, but it wasn't yours to kill. If you're casually putting up irrelevant (dirac) and WRONG files and still behaving as if you're entitled to browbeat everyone into qualifying themselves by them, when they are broken and need to be 'fixed', we can't proceed on that basis.
I fixed the mistake within 5 minutes! It was a very small one. The noise didn't expand over the full file length as it should. As such, segment 3 was exactly segment 5 (it had no noise).

BTW, that's what happens when you do things, and expose yourself to others. Mistakes. That's normal and it's rather impolite to use this as an anchor for discredit. You know it better. What about you posting real examples instead supporting your numerous wildly unproven claims? Anything? ("I ABX'ed unicorns, trust me")

That's left when the others run out of arguments. Personal attacks and indirect discredit. We've seen this more than often. It's ridiculous. You certainly can do better, eh?

All uploaded files are correct and up to date since about 20 hours, so please let's move on and talk about arguments.


Why can't I see any partials popping up in the examples 3 and 5?
Doesn't 3, 4 and 5 effectively dither?

Let's talk about this kind of stuff. Sorry for "killing a thread". But most arguments here are ridiculously out of scale, mostly learned by heart, simply dogmatic. Benefits on noisy signal straight unproven, or only valid in the case of very simple laboratory conditions. Only very few are interested in verification.

I don't accept what smells like bull****.

Again, insisting on adding -130dB of noise to an already given -60dB noise floor doesn't sound very intelligent. Can you explain why it would be intelligent? Maybe with 1-2 examples we can all verify? I explained and showed practically that my thoughts are correct. Most here, with a few exceptions, don't even matter, and prefer calling me things like a theorist, a party crasher.

Last edited by FabienTDR; 11th March 2017 at 05:09 PM..
Old 11th March 2017
  #411
@stinkyfingers: Please try a sat that doesn't add a DC offset.

Why null test? I doesn't say anything about audibility. Noise rarely nulls, but given the same noise properties (colour and distribution), they will sound equal and contain exactly the same information. "Hello, I am random, that's all I know". Noise is uncorrelated to anything surrounding it except for itself (the t= 0 case above), nulling is worthless in this case because the same sounding noise has unlimited amount of PCM shapes.

But when you explicitly add 0Hz content (in this case, it's synthesis), well, then you have noise + 0Hz content.

My point is that you cannot make noise produce any partials (well, except the 0Hz partial as you've shown!).

That's why we make sure to only truncate noisy regions (if we ever have to), as this keeps the spectral integrity intact.
Old 11th March 2017
  #412
Lives for gear
 
stinkyfingers's Avatar
 

Quote:
Originally Posted by FabienTDR View Post
@stinkyfingers: Please try without a DC offset. Feel free to post the analyzer plots and point me to one partial being generated. Noise never nulls. Noise is uncorrelated to anything surrounding it except for itself (the t= 0 case above).

But when you explicitly add 0Hz content, well, then you have noise + 0Hz content.

My point is that you cannot make noise produce any partials.
those two files are exactly the same source noise...
i generated a file of Gaussian noise. that is file "A"
then i rectified it. that is file "B".
it is the exact same file, just distorted.
if the only distortion was the DC as you claim, then removing the DC the files would null.
do they null ?
Old 11th March 2017
  #413
When you remove DC, both files contain exactly the same information (I am random), and sound equal. It just tried to filter both via a DC filter. And yes, they are equal. Same spectrum, same information.

What you are doing is 0Hz synthesis! You explicitly produced DC, this is a confusing, not really helpful edge case.

Of course it sounds different and all your gear will react differently in the presence of DC! You know that.


Try to truncate noise, just to stay on topic. Truncate whatever amount of LSB

Then show what significantly changed.

Last edited by FabienTDR; 12th March 2017 at 12:01 AM..
Old 11th March 2017
  #414
Lives for gear
 
stinkyfingers's Avatar
 

Quote:
Originally Posted by FabienTDR View Post
When you remove DC, both file contain exactly the same information (I am random), and sound equal. What you are doing is 0Hz synthesis! Of course it sounds different and all your gear will react differently in the presence of DC! You know that.
could you post the actual proof of those two files nulling with the DC removed because i can not get them to and would like to know how.
thanks,

(responding to your first bold statement, if they do not null, they are not equal. i know they are not equal because i can hear a difference, whether the DC is there or not. because there is a difference. there is more to a sound than an average/ static frequency spectrum plot, you know this. besides, i checked both spectrums out in multiple analyzers and they are far from exact. even at low fft size and smoothing you can see the difference.)

Last edited by stinkyfingers; 11th March 2017 at 07:49 PM..
Old 11th March 2017
  #415
Lives for gear
 
bogosort's Avatar
Fabien is correct: dither represents the minimal noise required to decorrelate the signal from quantization error.

Quantization error is the difference between the smallest step size (the value of the least-significant bit, or LSB) and the actual signal value. Since the quantized signal can only change in LSB-sized steps, the quantization error is correlated with the signal, i.e., the error depends on what the signal is doing. As everyone here knows, we can easily hear this when it happens, and it sounds wrong.

The difference between noise and distortion is that noise is uncorrelated with the signal, while distortion is correlated.

The way to get rid of quantization error (distortion) is to decorrelate it from the signal (turn it into noise). This is the purpose of dither. But if the LSB is already varying randomly -- i.e., if the quantization step size is smaller than the signal's noisefloor -- then the signal is self-dithering. Adding more dither noise is unnecessary.

There is no ambiguity here when it comes to conversion: a proper 24-bit ADC or DAC does not need dither, simply because the analog noisefloor is already providing 4 or more bits of dither. Likewise, there is no ambiguity when reducing word length to 16 bits or smaller; this should always be dithered, because we don't want to take any chances that the signal's noisefloor is below the step size.

Adding TPDF dither to a 24-bit signal is completely irrelevant in the vast majority of cases. However, adding noise-shaped dither, depending on final processing, might end up being harmful.
Old 11th March 2017
  #416
Nulling is radically inadequate to compare the audibility of signals. Do I have to debust all naive beliefs before we can discuss the central statements?!

Short:
With music signals and modern formats, explicit dither is grossly overrated.
In most cases completely ineffective.
The related anecdotes and personal stories are MUCH better explained via the well known perception biasses


You can have an almost infinite amount of representations for exactly the same sound in PCM. End of distracting nulling discussions.. ..I hope..

Noise is not "one" signal you could subtract from another. It's a statistical property saying: I don't correlate to anything surrounding me. As such, noise is radically different to say, tones, it follows very different laws. This property of randomness is very robust. Again, and this is the only reason why dither works in the first place, noise is not correlated to the main signal, it's free, independent, abusable and disposable in the purest sense.

Last edited by FabienTDR; 12th March 2017 at 12:03 AM..
Old 11th March 2017
  #417
Lives for gear
 
stinkyfingers's Avatar
 

Quote:
Originally Posted by FabienTDR View Post
Try to truncate noise, just to stay on topic. Truncate whatever amount of LSB.

Then show what significantly changed.
are you looking for a specific type of noise to be truncated, or can it be "whatever" as long as it is "full spectrum" (or close to)?

Quote:
Originally Posted by FabienTDR View Post
You can have an almost infinite amount of representations for exactly the same sound in PCM. End of distracting nulling discussions.. ..I hope..
you can Ethan Weiner yourself out of this one , but it is fact that noise has been (audibly) changed through distortion, which is what you have been preaching you can not do...
Old 11th March 2017
  #418
Yes, any wideband noise will be fine.
Old 11th March 2017
  #419
Gear Maniac
 
Yuri Korzunov's Avatar
 

Me seems, "truncate" is too general term.

"Rounding" show real math/physical meaning.
Old 11th March 2017
  #420
Truncation is fine imho: Remove (or overflow) N LSB, as it is done in fixed point processing, format conversion and AD/DA.

IMHO the meaning is clear. We talk about bits, not levels. You can't round bits.
Topic:
Post Reply

Welcome to the Gearslutz Pro Audio Community!

Registration benefits include:
  • The ability to reply to and create new discussions
  • Access to members-only giveaways & competitions
  • Interact with VIP industry experts in our guest Q&As
  • Access to members-only sub forum discussions
  • Access to members-only Chat Room
  • Get INSTANT ACCESS to the world's best private pro audio Classifieds for only USD $20/year
  • Promote your eBay auctions and Reverb.com listings for free
  • Remove this message!
You need an account to post a reply. Create a username and password below and an account will be created and your post entered.


 
 
Slide to join now Processing…
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Forum Jump
Forum Jump