Dithering 101

[Ethan Winer]

Originally Posted by bob katz

Ethan. Are you still hanging on that scientifically-fallacious premise that switching in the middle of a piece of music can be used to accurately assess sonic differences?

Bob, if you read that thread over in Lynn's forum you'll see this was a different test. We even raised the level way up during the final ending chord fade, and even then neither of us heard a difference.

I don't mean to beat this to death - well, okay, maybe I do - but I would like to hear from folks here how a test could be made to settle this for once and for all. A test that will let people here in the forum play sample files and "vote" for which they think is dithered versus truncated. A test that cannot be cheated. I do not accept that such as test is impossible, but I accept that folks seem to dislike all of my tests so far.

Folks, this is very simple - if truncating is so clearly audibly worse than dithering, then surely 100 DBT trials are not needed. Indeed, DBT trials are needed only to prove that something is not audible. Which is sort of my point here. If the advantage of dither is so obvious to so many, let's find a way to put it to the test for once and for all. I await everyone's suggestions.

--Ethan

[Ethan Winer]

Originally Posted by Cellotron

fwiw - I conducted a blind test on Brad Blackwood's forum a bit over a year ago. About 80% of the test takers (mostly ME's) were able to identify the truncated file out of about 8 other files that used different dithering algorithms.

Steve, is there a place I can download those files to check them out for myself?

--Ethan

[Ethan Winer]

Originally Posted by JohnRoberts

I'm not sure why Ethan takes such pleasure from stirring the pot

It's not pleasure as much as being dismayed to see people obsess over all the wrong stuff. I think of this as myth-busting more than stirring the pot.

I'm also still waiting for lukejs to explain more about what he heard, because I'll probably have some follow-up comments and questions for him. You still there Luke?

--Ethan

[Cellotron]

Originally Posted by Ethan Winer

Steve, is there a place I can download those files to check them out for myself?

--Ethan

No - I removed it from my site about a year ago. I'll see if I can locate a backup sometime to repost it if I can - but really don't have anytime to do this currently.

You can still see the original discussion on the dither shootout at PSW Recording Forums: Brad Blackwood => NEW DITHER SHOOTOUT (#1) NOW POSTED!!

Best regards,
Steve Berson

[johnlink]

Originally Posted by masteringhouse

Spectrafoo has no effect on the audio. John what you are hearing is the effect of the harmonic distortion created via truncation. The additional overtones make the audio sound a bit higher in pitch and volume, but the fundamental frequency and levels were the same in all samples.

Just be to clear about my perceptions, here is a summary:

Truncated file: same pitch and volume as undithered file, lots of distortion

Dithered file: higher pitch than undithered file, lower volume than undithered file, much less distortion than truncated file


I still do not understand why the dithered file sounds so quiet relative to the other two files.

Are you saying that the increase in the apparent pitch of the dithered file is due to the dithering? Although I hear much less distortion (i.e., noise) in the dithered file than in the truncated file, I certainly find the apparent shift of pitch to be a serious distortion in its own right.

John Link

[masteringhouse]

Originally Posted by johnlink

Just be to clear about my perceptions, here is a summary:

Truncated file: same pitch and volume as undithered file, lots of distortion

Dithered file: higher pitch than undithered file, lower volume than undithered file, much less distortion than truncated file


I still do not understand why the dithered file sounds so quiet relative to the other two files.

Are you saying that the increase in the apparent pitch of the dithered file is due to the dithering? Although I hear much less distortion (i.e., noise) in the dithered file than in the truncated file, I certainly find the apparent shift of pitch to be a serious distortion in its own right.

John Link

By undithered file I assume that you're taking about the 24 bit source.

All 3 files are 500Hz tones, there is no pitch shift that I hear, just the additional harmonics in the truncated version. The dithered one does sound a bit lower in volume. I normalized all of these to -14dBFS peak, the RMS value of the dithered version ends up being lower. Not exactly sure why this is the case, but I'm thinking this may have something to do with the noise shaping in the L2. I suppose that they could be raised to the same RMS level, but wanted to keep the processing simple. Anyway that wasn't the purpose of the samples, but merely to demonstrate the harmonic distortion inherent in truncation and what to listen for when comparing dithered to non-dithered 16 bit files.

[johnlink]

Originally Posted by masteringhouse

By undithered file I assume that you're taking about the 24 bit source.

That's correct. I will refer to it from now on as the source file.

All 3 files are 500Hz tones, there is no pitch shift that I hear, just the additional harmonics in the truncated version.

I find it very interesting that you don't hear the pitch shift. I wasn't expecting it, and I find it as obvious as can be.

The dithered one does sound a bit lower in volume. I normalized all of these to -14dBFS peak, the RMS value of the dithered version ends up being lower.

That's what I'm hearing.

Not exactly sure why this is the case, but I'm thinking this may have something to do with the noise shaping in the L2. I suppose that they could be raised to the same RMS level, but wanted to keep the processing simple. Anyway that wasn't the purpose of the samples, but merely to demonstrate the harmonic distortion inherent in truncation and what to listen for when comparing dithered to non-dithered 16 bit files.

I think it is important to compare files that are the same volume. Would you post files that you don't normailze? I think that would allow a more relevant comparison.

[masteringhouse]

Hi John.

I ran the downloaded files into another FFT analyzer and they all showed the same fundamental at 500 Hz. Please try this on your end to verify. I posted this stuff last week and have since deleted the session so I would have to re-create in order to get the files at their original level. I'm not sure that it will prove anything that isn't obvious in these though.

Let me know if you come up with anything regarding a pitch shift using an analyzer on your end, that would truly be interesting ...

[hummer]

apologies if i re-hash points that have already been made...

after reading this thread, i think i fully understand what ethan is trying to say... that the difference between truncation and dithering *to 16 bits* is inaudible under most listening conditions.

while i am inclined to *more-or-less* agree with this statement, (i don't think i'd be able to tell the difference when listening to loud rock music on my ipod) what i don't understand is why anyone would bother arguing this point at all. i find it irresponsible, misleading, and unprofessional. especially coming from an expert (or, someone in the audio industry)

people come to this forum to learn. i understand that this *discovery* that you've made may have some merit to it, but why would you respond to the original poster's question with this kind of information?

i mean, yeah... congratulations. you have discovered that the success of digital audio is a combination of subtle elements.

it is true that distortion components might not be immediately apparent, especially to the untrained ear. but the "edginess" of digital audio was overcome by the implementation of better practices like dither, among others (oversampling, better filter design, higher precision calculations etc). these "small" improvements account for the high quality audio we are now able to produce. you may be able to remove one of these elements without an immediately obvious penalty, but you will slowly approach the stone-age of digital.

the problem with your statement is that dither IS necessary if one is to achieve distortion-free bit reduction of digital audio.

whether or not distortion is evident in your experiment is completely irrelevant, because it only "proves" that it doesn't matter in one specific scenario (or at best, in the case of 16 bit truncation of musical program material, where the average level is high enough to mask the quantization distortion). what if (for some crazy reason) the audio were to peak at -30 dbFS, and average around -45? surely, dither vs. truncation would be more apparent. or, what about 12-bit audio? (where the difference between truncation and dithering is obvious) yes, i'm using extremes here, but i'm trying to point out that your observations are scenario-dependent.

dither IS utilized within DAWs and various signal processors, at various bit depths, in order to avoid successive truncation and culmination of quantization distortion artifacts. digital mixers and properly designed fade editors employ re-dithering. as do many processors, as even relatively basic DSP calculations will expand internal word length. (i'm sure you know this stuff...)

your words are fine for a layman's debate, but this is a forum of professionals, and in particular, a forum of detail-minded, quality-oriented professionals that pride themselves on understanding the nuances of digital audio.

there is an interesting experiment that i recommend. using a processor that can reduce bit depth, start with a 24 bit source, and reduce the word length to something silly, like 4 or 8 bits, without dither. now, increase the word length by 1 or 2 bits every 10 seconds or so. when you reach 16 bits, you will find that you can hear that the expansion of dynamic range could continue. and it will! after a few times through this experiment, it becomes increasingly easy to appreciate the increase in dynamic range, and lack of distortion elements. it's like a breath of fresh air.

i can hear the difference between 16 bit-truncated and 24 bit material in a studio environment. therefore, i can hear the difference between 144 dB of dynamic range, and 98 dB. therefore, elements of the audio in excess of 90 dB below full scale ARE important! dynamic range IS critical. not to mention, the more the better. flat re-dithering increases the perceived dynamic range of 16 bit audio considerably (20 dB or more, IIRC). and of course, distortion is eliminated.

like i said, i realize that you aren't trying to disprove the science of quantization, or the FACT that dither eliminates distortion. you are only saying that in your audio example, the difference between truncation and dither could not be observed. and i believe you.

what i am telling you is that it's irresponsible of you to propagate this kind of information, especially to a curious newbie looking for education.

[johnlink]

Originally Posted by masteringhouse

Hi John.

I ran the downloaded files into another FFT analyzer and they all showed the same fundamental at 500 Hz. Please try this on your end to verify. I posted this stuff last week and have since deleted the session so I would have to re-create in order to get the files at their original level. I'm not sure that it will prove anything that isn't obvious in these though.

Let me know if you come up with anything regarding a pitch shift using an analyzer on your end, that would truly be interesting ...

I will investigate further and report my results.

John Link

[Francis Vaughan]

I'm very much an outsider here, but felt it worth making a point or two. Probably blindingly obvious, but none the less possibly missed. (There have been a few statements about dither that were either technically just plain wrong, or missed the point of the discussion.)

But a reminder. If you have a 24 bit source that has a noise floor worse than 96dB below the 0dB mark, and truncate it, your truncated signal is actually already dithered. Not just noisy, but it will have all the lower than LSB resolution capabilities of noise shaped dither, simply defined by the spectrum of the noise floor. It matters not what the source of the noise is. Additive noise from multiple analog sources via their DA and thence mixed in will do it.

So, any test for the value of dither or not during conversion of 24 to 16 bit needs to ensure that the 24 bit source has a noise floor below the truncation level, at all frequencies. I would suspect that this is harder to do than many would hope.

As to the early question, dither once and last? Absolutely. The key thing is that dither encodes information into the the LSB that resolves information at levels lower than the LSB in a restricted range of frequencies (dependent upon the dither profile, and thus chosen for best effect with the recorded material.) This encoding vanishes totally if anything is done to the data stream ever again. It will simply turn into a raised noise floor. The special case of taking a 24 bit source, translating to 16 for some reason and then injecting back into 24 bit space is the one time where dither needs to be applied mid chain - the first 24 to 16 bit conversion should be dithered.

[Bobro]

Originally Posted by Francis Vaughan

As to the early question, dither once and last? Absolutely.

But in DAW DSP there can be a huge number of dithering or truncation instances in a mix already, any time the processing is carried out at a higher bit rate than the project rate by a plugin or whatever. Not only that, but there's another kind of "dither" that might be happening in a number of plugins, "denormalization".

Pentium IV processors apparently slow way down in certain circumstances which appear in DSP and the way around it, for quicker and realtime performance, is to add some kind of low-level noise. Don't know how it's done, but I do know that, for example, rendering a file in Csound offline, the rendering time went from 2:48 (m:s) to 1:16 when I inserted the "denorm" opcode. I'd say it's a safe bet that some VST softsynth plugins for example do this.

Lessee, googling..." VST denormalize plugin Pentium IV"... yip it's an issue and discussed by the developers.

What kinds of dithering do and don't get done in the myriad of plugins, who know?

-Cameron Bobro

[bob katz]

Originally Posted by Francis Vaughan

I'm very much an outsider here, but felt it worth making a point or two. Probably blindingly obvious, but none the less possibly missed. (There have been a few statements about dither that were either technically just plain wrong, or missed the point of the discussion.)

But a reminder. If you have a 24 bit source that has a noise floor worse than 96dB below the 0dB mark, and truncate it, your truncated signal is actually already dithered.

This is true If and ONLY if there is no low level information BELOW the RMS noise floor! People tend to forget that we DO hear signal BELOW the noise. There's plenty of life below the noise floor!

To be really scientific about it, before truncating, you would have to listen to the LSBs below the 16th bit (isolated) to make sure there is no modulation information in there. If you hear modulation (even if the LSBs sound distorted), it is likely contributing (subtly) to the ambience of the source and thus should be redithered to 16.

On a related issue.... people sometimes think that if you start with a 16 bit source and further mix it or process it, since it was "already dithered", that you don't have to dither it again to get it back to 16. This is entirely INCORRECT.

BK

[Francis Vaughan]

Agreed, indeed there is duality in the mathematics of how we hear below the noise floor and dither's operation.

My point was being made for particular purpose. The point being that if you compare a truncated 24 bit signal with a dithered signal, if the original signal had stochastic noise of the order of -96dB the truncated signal would already have had at least part of the benefit of a simple dithering. Dithering that still would need to be applied to a signal that had a superior noise floor to avoid quantisation noise being a clearly audible issue.

There is an implicit challenge here. If you have a 24 bit signal and claim that truncation sounds no different to a dithered conversion, you need to first ensure that you not not already have an implicitly dithered source. At least dithered enough to have decorrelated the quantisation noise. There is no doubt proper dithering can do better, and depending upon choice of function get a good 18 bits worth of resolution in chosen parts of the spectrum. But just adding some shaped noise, even if by accident before truncation can get a long way there. Again there is a duality here. After all the operation of dither in a ADC can be the true addition of noise to the analog signal prior to the quantiser. Indeed a mic pre can be easily noisy enough to do the job with no further effort.

A further point, it is worthwhile distinguishing between the sonic advantages of shaped dither's ability to allow resolution below the LSB, and the clear necessity to decorrelate quantisation noise. The former is a nice bonus, and adds to the sonics, the latter a grim necessity to avoid clearly sub-optimal results. That both come in the same package is simply a neat result.

[UnderTow]

Francis,

Everything I've learned (from people much more knowledgeable than me) says you are wrong on both counts.

You need noise of a particular distribution to dither properly (not just any noise). And you need to dither every bit reduction. Not just the final bit reduction.

Alistair

[johnlink]

I listened to the three files posted by Tom Volpicelli (masteringhouse) and today I find that the dithered file seems to have the same pitch as the source file. I'm glad that that is the case, but I do wonder why the truncated file sounded a half step higher yesterday. The application Amadeus confirms that the truncated file has a fundamental of 500 Hz.

I find that the dithered file needs to be boosted by 8 db in order to be as loud as the source file. With that boost there is audible distortion, but much less than the truncated file.

John Link

[Francis Vaughan]

I'm probably not making myself clear enough. There is a difference between simple dithering, and "properly" dithering. But the point I'm making (badly it seems) is that the addition of simple stochastic (AIWN) noise will achieve the bulk of the effect. There are many ways in which it can be improved upon, simple tricks like shaping of the noise spectrum are a start, but only a small one. Eventually you get into the realm of much more sophisticated algorithms that can perform better again, and can be crafted to make best knowledge of the aural nature of the signal.

But again, that isn't what I was trying to talk about. It was, again, to point out that the addition of a suitable quantity of stochastic noise before truncation will provide a very large fraction of the effect of the dither process, and will be enough to cause decorrelation of the quantisation artefacts. The whole point was to outline a flaw in simple "dither is useless" experiments. Not to suggest that you can simply add a bit of noise and be happy.

The issue with when and when not to dither is becoming more complex than I had intended the reasoning to be about. Also I was not making myself clear again. Sorry. I did mean that dither should be done on each bit reduction. I was making trivialising assumptions about the workflow (i.e. that once in the 24 bit domain you would keep it there.) I was addressing a sometimes heard confusion that dither needs adding at other steps.

[UnderTow]

Francis, thanks fro clearing that up.

Alistair

[masteringhouse]

Originally Posted by johnlink

I listened to the three files posted by Tom Volpicelli (masteringhouse) and today I find that the dithered file seems to have the same pitch as the source file. I'm glad that that is the case, but I do wonder why the truncated file sounded a half step higher yesterday.

John -

It doesn't sound as if it there is a pitch change today? If this is the case, I'm wondering if there are clocking issues with your converter or the file format? The files have been unchanged since I first posted them. Or maybe it's a psychoacoustic effect from it being at a different level, and the meds kicked in today ;-)

[UnderTow]

Originally Posted by masteringhouse

John -

It doesn't sound as if it there is a pitch change today? If this is the case, I'm wondering if there are clocking issues with your converter or the file format? The files have been unchanged since I first posted them. Or maybe it's a psychoacoustic effect from it being at a different level, and the meds kicked in today ;-)

I suspect it was a psychoacoustic effect due to the high pitched shaped dither noise.

Alistair

[masteringhouse]

Originally Posted by UnderTow

I suspect it was a psychoacoustic effect due to the high pitched shaped dither noise.

Alistair

Given that the purpose of this type of noise shaping is to stay out of the audible range I would find that odd, or do I need more meds? If so, and more apropos this thread, how would this affect our perception of dithered versus non-dithered audio?

[Ethan Winer]

Originally Posted by hummer

what ethan is trying to say... that the difference between truncation and dithering *to 16 bits* is inaudible under most listening conditions.

Exactly.

people come to this forum to learn.

You don't think learning what does and does not matter in practice (versus theory) is useful information for newbies? This is as much about consumerism and value as pro audio IMO. What I find irresponsible is the dozens of magazine articles over the past 30 years spreading false information. Not just about the importance of dither, but many other untruths too. I'll stop short of saying they lie to help their advertisers sell gear, because I think the writers are just as much victims of these myths as their readers. But I blame them because they're the ones who perpetuate the myths.

whether or not distortion is evident in your experiment is completely irrelevant, because it only "proves" that it doesn't matter in one specific scenario ... what about 12-bit audio?

If you read this thread you'll see I have asked repeatedly for the pro-dither people to post any pair of files they have showing dither being audible at normal listening levels. If you think my test is not representative of all music, I ask you - no, I beg you - to post a pair of files where the effect of dithering can be heard on normal material at normal listening levels. I have also stated more than once that I'm talking only about going from 24 bits down to 16. For example HERE.

i realize that you aren't trying to disprove the science of quantization, or the FACT that dither eliminates distortion. you are only saying that in your audio example, the difference between truncation and dither could not be observed. and i believe you.

Then why are you telling me ?

--Ethan

[UnderTow]

Originally Posted by masteringhouse

Given that the purpose of this type of noise shaping is to stay out of the audible range I would find that odd, or do I need more meds?

They are certainly not out of the audible range! They are in a range to which we are less sensitive to. That is not the same thing. (Fletcher-Munson curves).

Anyway the dither in your file is not at typical levels by a long shot. The noise shaped dither in your example is nearly as loud as the 500Hz tone! It is only 10 dB down and is painfully audible to my ears.

If so, and more apropos this thread, how would this affect our perception of dithered versus non-dithered audio?

Again, these are not typical levels but in general I am not a fan of these ultra shaped dither. Maybe because they add a sharpness to the signal? I haven't done any proper tests to determine if this is the case. I just tend to choose other dither types.

Alistair

[Storyville]

Originally Posted by Ben B

Personally, what I find difficult to comprehend is how dither can be of much (if any) benefit when the music is loud. I too have heard the "fitzing and sputtering" (great descriptives, by the way) of quantization noise at the very end of fades, but the rest of the time -- when the music is loud -- does it have any benefit? It's a rare circumstance in recorded popular music that a reverb tail decaying to -96 dB would be fully "exposed" (and not modulated upwards by other simultaneous mix elements at higher levels) for dithering to have its intended benefit in any location other than the very end of a song. While I do see the benefit it has at the ends of fades, I'd love to hear a clear and scientifically valid description of how a word-length-reduced recording benefits from dither when the music is loud.

Anyone have the answer to this?

I honestly don't know....but logically I would think dither is useless when the music is at full volume. Of course, its supposed to be very hard to notice when the music is at full volume. But, you can always automate a dither plug in to come in at some point while the record fades out.

[masteringhouse]

Originally Posted by UnderTow

They are certainly not out of the audible range! They are in a range to which we are less sensitive to. That is not the same thing. (Fletcher-Munson curves).

Understood. I should have stated it that way.

Originally Posted by UnderTow

Anyway the dither in your file is not at typical levels by a long shot. The noise shaped dither in your example is nearly as loud as the 500Hz tone! It is only 10 dB down and is painfully audible to my ears.

True indeed. Not sure if this would be the cause for a pitch shift though. I've always thought that difference tones were spaced closer together than the fundamental and dither present in this example.