Dither - My Ears can't hear it

[Ethan Winer]

Hash? Where are you, Amsterdam????

Heh.

I hold dither above the others because IME there does seem to be a "signal" there

Yes, there's a signal there all right, but masking makes it inaudible. Like the example I think I mentioned earlier of the piano tracks posted at Lynn Fuston's forum where the acoustic noise floor was around -70. Guys insisted they could hear the dither, even though it was 20 dB below the room noise. This is a huge factor! How many tracks has anyone here produced where the noise floor was anywhere close to -90 where dither operates?

I've put a test method on the table

Did I miss that? Which post number? Yes, please, let's do it! But the test must be such that people can't cheat. This is the main problem I see because people can subtract one file from another and isolate just the dither. And from that they can tell which of two test files is dithered and which is not with one more subtraction.

This is why I made my test file the way I did. If dither is switched on and off in the middle of a passage, and nobody can even tell where a switch occurs, I maintain that proves the dither was inaudible. Maybe only for that one music example and one dither flavor, but proven inaudible none the less.

I will side with you IF the results tell me to as long as you can say the same for me .

Hell, I would love to be proven wrong. Go for it guys!

--Ethan

 

[NL5]

I can tell right away if my tracks are dithered or not. I'll put up some files later today with different dithering, and no dithering and see if anyone can tell which is which.

Ok, to Ethan's credit, I probably overstated this somewhat. I am pretty sure I hear a difference. Maybe I do, maybe not. It is fairly subtle, but everything we do as engineers is SUBTLE changes. It's all the subtleties that add up. Besides, dither isn't costing anything, right?!??!!?

Anyway, I posted a thread with a quick listeneng test if anybody wants to try it.

https://homerecording.com/bbs/showthread.php?p=2849006#post2849006

 

[pipelineaudio]

that test project I posted earler will certainly let you hear the difference between on and off at 9 bit, but its a ridiculously overblown scenario, where I took a cymbal tail that was originally peaking at -70 and cranked it to the moon

 

[SouthSIDE Glen]

Yes, there's a signal there all right, but masking makes it inaudible.

I was talking metaphorically - hence my use of the term "metaphorically" twice in one sentence . By "signal" I was referring to notciable results above random chance.

Not everything that's louder masks everything that's softer. The single idea that dither takes place "below" the noise floor is in and of itself no guarantee of masking.

Did I miss that? Which post number? Yes, please, let's do it! But the test must be such that people can't cheat. This is the main problem I see because people can subtract one file from another and isolate just the dither. And from that they can tell which of two test files is dithered and which is not with one more subtraction.

That's where I was talking baout the double-blind testing. OK, I didn't go into a lot of specific details, but the broad strokes were presented.

And yes, I agree, that "cheating" needs to be addressed as well as possible. Methods like yours will work well, plus I have a few other protections in mind. It'll take some time before I can get it completly up and running as I'm finishing up a couple of other projects on the site right now that I GOT to get running before I start anything else, but in the meantime I can collect the files and info from everybody and start on the basic test design.

G.

 

[masteringhouse]

The page below demonstrates the effects from 16 bit truncation and dithering:

http://www.masteringhouse.com/dither/dither.html

It contains graphs from SpectraFoo as well as providing audio clips in 24 bit format. You can download these and compare in your favorite DAW. To most I feel the difference is very audible.

While I suppose some may argue that the initial level of the signal was so low that the distortion would be inaudible in normal circumstances, in earbuds, mp3s, etc., one has to consider cumulative effects of this from processing, raising gain, summing multiple tracks, etc. At any rate, I don't see why any good engineer would want this distortion present at any degree or to "sweep it under the rug" when it can be improved upon so easily. Isn't that one of the things good engineering is about?

Best,
Tom

 

[pezking]

Ultra-high sample rates are BS because nobody can hear past 22 KHz. Phase shift is BS because it's inaudible.

The reasons for high sample rates are for better/easier engineering - not necessarily an increase in bandwidth for listening purposes. What phase shift do you refer to? But like I said let's stay on topic! (I'd be happy to discuss these points in another thread! )

I don't see why any good engineer would want this distortion present at any degree or to "sweep it under the rug" when it can be improved upon so easily. Isn't that one of the things good engineering is about?

Very good points. Low level noise is far preferable to distortion.

Like Ethan said already - great discussion (not argument). Well done to mattkw80 for starting a very interesting debate!

 

[rayc]

In Cakewalk Pro Audio 9.3 the rendering to 16 from 24 isn't done, as far as I've been able to discover, in or after the mixdown. The program requires that the entire project be rendered at one time prior to mixdown.
This had me spun - all the effort, money & time to do 24 and it's batch converted to 16 (irreversibly)!
I worked around it by mixing at 24 then importing the mixed stereo pair into a temp project & rendering it to 16 there - BUT I had some problems with clicks, pops, etc in a hot mix - I didn't mind reducing levels in the mix to get a cleaner result - I mean I ought not have had them so high in the 1st instance but I was just experimenting with software faux mastering.
Any clues as to a better way around (in cakealk pr audio that is)?

 

[masteringhouse]

Very good points. Low level noise is far preferable to distortion.

Absolutely, and decorrelated noise is preferable to noise and distortion that is signal dependent. The psychological principle of habituation essentially states that we become less sensitive with repetition of a stimulus. If the noise is consistent, we basically psychologically mask it out. It's when it changes with the signal (correlated) that we take more notice.

Also shaping the noise so that it is in a region that is less audible is also usually preferred over broadband noise, even if it is consistent and at a low level.

 

[pezking]

Also shaping the noise so that it is in a region that is less audible is also usually preferred over broadband noise, even if it is consistent and at a low level.

Which is where high sample rates can help, because you can make that dither noise largely supersonic

 

[masteringhouse]

Which is where high sample rates can help, because you can make that dither noise largely supersonic

Well ...

Ultimately it needs to be converted to 44/16 and dithered at that point.

We'll save the use of high sample rates for another interesting debate

 

[pezking]

Well ...

Ultimately it needs to be converted to 44/16 and dithered at that point.

We'll save the use of high sample rates for another interesting debate

Unless you're mastering to DVD, or something similar

But yeah, back to dither...

 

[SouthSIDE Glen]

Which is where high sample rates can help, because you can make that dither noise largely supersonic

Not picking on you here, pez, because I've liked your responses thus far; just using the quote as a re-entry point.

What many folks forget - or don't understand to begin with - is that when one is talking about dither "noise", that is meant more in the definition of information theory noise, not "noise" as we are used to thinking about it in the audio world.

To think of adding "noise" to the signal in the form of a very low volume hiss or static, and that it's just plain not audible because it is so low in volume, or changing it's frequency to make it less audible, is really not quite thinking about it the right way. While if one took just a pure dither signal and amplified it's volume, it probably would wind up sounding like a type of noise, that's a bit misleading of a consequence. We shouldn't think of dither as an analog noise.

In the case of truncation dithering, it's digital noise being added to mask digital information. "Noise" in this case does not specifically mean sound, it simply means a lack of information. Truncation alone causes a falsification of information that follows a pattern, much (but not exactly or specifically) like clipping falsifies peak information following a regular pattern. This pattern of truncation is (barely and subtly) audible simply because it IS a pattern, i.e. it carries it's own information. Dither is the removal of that pattern and the removal of the truncation information, removing the false information artifacts of truncation.

In this way it can be said that dithering itself does not have a sound, but rather it is the removal of a sound. This is why dithering is so content-dependant, and why dithering itself is hard to pin down as having it's own identifiable sound.

Those not steeped in information theory can find this whole concept very hard to understand. Noise is noise, right? Audible is audible, right? Not necessarily.

G.

 

[Ethan Winer]

Ok, to Ethan's credit, I probably overstated this somewhat.

Yeah, just a little.

I posted a thread with a quick listeneng test if anybody wants to try it.

IMO, as soon as you use lossy compression, even at a high bit-rate, that will change the sound enough to render the dither comparison moot.

--Ethan

 

[Ethan Winer]

The page below demonstrates the effects from 16 bit truncation and dithering

Yes, but the differences are all 105 dB down or even lower. So we're already at least 10 dB below the noise floor which in my estimation is inaudible.

Note that I never said dither couldn't be measured, or that it's not better than truncated in theory. Only that it can't be heard in practice.

--Ethan

 

[Ironklad Audio]

ok...without getting dick-deep into your guy's argument, i can tell you that if i slap a plugin on the 2-buss of a mix and engage the dither, i can hear a very distinct and audible difference when changing the type of dithering curve that is being applied...now admittedly, i've never tried just turning the dither on/off and listening for a difference, but the fact that there is an obvious difference between the different dithering algorithms would lead me to believe that the dithering itself is audible.

and when it comes down to it...what does it really matter? how much extra work does it take to click "dither"? i can see the point in arguing the necessity of a $150 boutique cable versus a $2 zipcord, but this really just seems silly...if you think the dithering sounds better, click on the dither button. if you don't, then, well, don't.

also, am i the only one who's waiting for ethan to say something like "the only way you can properly hear the dithering is by adding more bass traps"?

 

[masteringhouse]

Yes, but the differences are all 105 dB down or even lower. So we're already at least 10 dB below the noise floor which in my estimation is inaudible.

Note that I never said dither couldn't be measured, or that it's not better than truncated in theory. Only that it can't be heard in practice.

--Ethan

As usual it depends. Take distortion at that lower level, sum 24 tracks of it, process it, limit, raise the gain to today's levels and you pretty much have what is demonstrated in the audio samples.

ok...without getting dick-deep into your guy's argument, i can tell you that if i slap a plugin on the 2-buss of a mix and engage the dither, i can hear a very distinct and audible difference when changing the type of dithering curve that is being applied...now admittedly, i've never tried just turning the dither on/off and listening for a difference, but the fact that there is an obvious difference between the different dithering algorithms would lead me to believe that the dithering itself is audible.

Same here. I can hear the differences between different types of noise shaping for example. I'm not trying to be "Mr. Golden ears" here, but I can and apparently so can some of my clients when working here. Can I identify which noise shaping is being used? No, but I don't particularly care to. Turn it on and hear which sounds best for the material at hand.

and when it comes down to it...what does it really matter? how much extra work does it take to click "dither"? i can see the point in arguing the necessity of a $150 boutique cable versus a $2 zipcord, but this really just seems silly...if you think the dithering sounds better, click on the dither button. if you don't, then, well, don't.

Absolutely. Dither is a no-brainer assuming that you use it correctly.

also, am i the only one who's waiting for ethan to say something like "the only way you can properly hear the dithering is by adding more bass traps"?

I wonder if anyone can listen to a track and identify if bass traps are or are not being used in a room? Irregardless it doesn't mean that they aren't a useful tool when needed, or should we mask out this effect with bass synth?

 

[SouthSIDE Glen]

Yes, but the differences are all 105 dB down or even lower. So we're already at least 10 dB below the noise floor which in my estimation is inaudible.

This is a big part of the misunderstanding, Ethan. It's not "below" anything; it's applied to what is there. The idea that because the analog noise level is at a louder level than the dithering that any change made by dithering is masked by noise is untrue. Also, as I explained in my last post, dithering itself is not noise in the typical analog sense, it is the forced loss of digital information through a "smearing" or randomization of that information.

Truncation causes an identifiable, repeatable and repeated pattern of distortion. It's "sound" is based upon that pattern (in info theory jargon, that pattern is extra information that is being conveyed). This is true regardless of whether it's a pristine, noiseless waveform, or whether there is circuit or ambient noise modulating it as well. Either way, truncation artifacting does appear as extra information encoded within the signal and the noise (they are one waveform, the noise really is part of the signal at that point).

There is no "masking", because the information affects the entire waveform, and also because truncation patterns are *not* noise in the analog sense; they're not even noise in the digital IT sense. Pink or white noise, even at higher volumes, does not effectively mask truncation artifacting because they are different animals with different audible effect. No more than a cello will effectively mask a saxaphone, even if the cello is louder.

It's equally true OTOH that - regardless of analog S/N - ditering is a method of removing that information and it's resulting perceived sound from both the digital signal and the digital representation of the analog noise. The perceived sound may be subtle, yes, and it's removal even subtler (it's harder for the ear to hear what is missing than what is there), but that's because of the nature of the beast itself, not because any masking from analog noise sources makes it inaudible.

G.

 

[mshilarious]

Yes, but the differences are all 105 dB down or even lower. So we're already at least 10 dB below the noise floor which in my estimation is inaudible.

Note that I never said dither couldn't be measured, or that it's not better than truncated in theory. Only that it can't be heard in practice.

--Ethan

Sorry Ethan, but I can't fully agree. I don't disagree that the effects of dither can be difficult or impossible to hear on much program material. But I can't agree that the effects cannot be measured and shown to be potentially audible.

As a circuit designer I'm not overly fond of listening tests, because it's common for listeners to select material that is technically inferior. That can be a design decision, but it should really be disclosed in that manner.

If you told me that you had a technique that could reduce THD to unmeasurable levels at little cost, I would generally take it. We aren't talking about $7,000 speaker cables, we are discussing a tool whose need is scientifically demonstrable, and which comes bundled with most audio software.

The benefits of dither are most audible on quiet material with low noise floors. That's obvious. So it's much more indicated on classical music than rock. Your stipulation that the material not be amplified beyond normal acoustic levels is telling. This is because the ear grows insensitive to distortion at low levels. But as a recordist, I can't presume that the listener will calibrate their monitors appropriately and never crank it.

Having said that, here is a simple test file: a 1kHz sine wave at -47dBFS RMS and pink noise at -74dBFS RMS. Roughly speaking, that should correspond to an acoustic noise floor around 17dBA or so, and a source at 49dBSPL. So, the quiet passage of the aforementioned flute, perhaps. Not a typical situation for a rock band, but not beyond the pale for classical music. Although your cello probably has enough overtones of its own to mask the effect . . .

http://www.naiant.com/images/1kHz_with_PN_trunc.wav

The result of truncation is a number of distortion products above the noise floor, totaling around 0.2% THD. That's not a huge figure, it's very difficult to hear (especially if the signal playback is calibrated--but if it's turned up, it's much more audible). THD will rise as the signal level drops, however. With the source at 40dBSPL, THD is 1%.

(As an aside, you made a conceptual error when dismissing masterhouse's FFT analysis above. A harmonic at -105dBFS is not made inaudible by a -74dBFS RMS noise floor, which analysis of my file will show)

Again, what is the cost to eliminate quantization distortion? It's not like I can buy a copy of Wavelab without the UV22HR bundled with it. To me, it's simply good engineering practice.

 

[masteringhouse]

Excellent point! I shoulda mentioned that as well. Also analog noise of the converter not considered in the equation, it's not 10 db below the noise floor. I think SpectraFoo's graphic for scaling is not completely accurate.

 

[mshilarious]

Here's an irony: I just finished a track that I left undithered, because for some reason the Apogee plug flipped out and generated ridiculous noise at the fadein from black to room noise before the first note started Further investigation of that phenomenon is required