Dithering 101

[Ethan Winer]

Originally Posted by peeder

Oh and btw Ethan I bet I could write code to look at the distribution of your LSB and find where your dither was applied and not.

Perhaps, but that's cheating and violates the spirit of the test. I never said dither can't be measured, nor would I claim that for stoopidly high sample rates or bit depths either. We can measure to orders of magnitude beyond what is audible, and I'm addressing only what can be heard and thus actually matters.

--Ethan

[twwalsh]

Originally Posted by Ethan Winer

It depends on how you define loud. If you play a track with the volume as loud as you can reasonably stand, and cannot hear when dither is used or not, then dither doesn't matter. If you have to turn up the volume even more than that to tell, then Yes it's audible in that one contrived situation but still doesn't matter in practice.

Rather than argue, I urge you to download my test file linked earlier. Its low recorded level is as favorable to dither as I could make it, so let's see if you can identify which places use dither. Then we'll both know for sure.

Deal? And not just you, but everyone else here who believes dither matters.

--Ethan

when i have time i will download your test.

in my own work, however, i can hear the 16 bit dither in voxengo elephant IMMEDIATELY and unmistakably when i turn it on...listening at 80-85 db to pop/rock.

[BushmasterM4]

But has anyone tested running two instances of dithering ? And are all created equal. Who decided what the freq. of the noise should be ? Do all dithering methods use the same freq. ? Did they base it on visual aids or thier ears ? Was it created back when digital recording was new and had bugs ? I asked the original question thinking there was a definate answere. I guess there isnt ? :) I need to do some experiments on the subject. I really appreciate all the input and it makes for good reading.

[Coyoteous]

- quick aside: flat TPDF for dither to 24 bit?

[Lagerfeldt]

Yes

[jamsmith]

Whew, I can't wait to get my lab set up to test some of these things for myself. Back when digital was fairly new, people did tests on 4-bit signals and the effect of dithering was phenomenal.

I think it was asked about the need for dithering for loud passages? Well, no, but then a tiny bit of noise wouldn't hurt either. Plus, lot of loud music has stops where you might hear a cymbal decay before the barrage starts up again. Doesn't it make sense to have dither on for those?

"Dither noise should be within one quantization step of the LSB." That is correct. Dither is noise, why would you want to add more noise than you need?

[BushmasterM4]

Originally Posted by jamsmith

Whew, I can't wait to get my lab set up to test some of these things for myself. Back when digital was fairly new, people did tests on 4-bit signals and the effect of dithering was phenomenal.

I think it was asked about the need for dithering for loud passages? Well, no, but then a tiny bit of noise wouldn't hurt either. Plus, lot of loud music has stops where you might hear a cymbal decay before the barrage starts up again. Doesn't it make sense to have dither on for those?

"Dither noise should be within one quantization step of the LSB." That is correct. Dither is noise, why would you want to add more noise than you need?

Dont need more. At times I have plugins that incorporate dithering and my Recording software has a dithering option. Ive been using two at times not realizing it. Just wondered if its a big no, no. Tests will tell.

[Ethan Winer]

Originally Posted by twwalsh

when i have time i will download your test.

Please do, and report what you find in this thread. The silence here and at Lynn's forum from those claiming that dither is critically important is deafening.

in my own work, however, i can hear the 16 bit dither in voxengo elephant IMMEDIATELY and unmistakably when i turn it on...listening at 80-85 db to pop/rock.

If this is really true and a change is blatant, I can only imagine that something else unrelated is going on.

--Ethan

[bob katz]

Originally Posted by Ethan Winer

Please do, and report what you find in this thread. The silence here and at Lynn's forum from those claiming that dither is critically important is deafening.



If this is really true and a change is blatant, I can only imagine that something else unrelated is going on.

--Ethan

OK, Ethan. Dither is NEVER that obvious until you discover later how you screwed up a master. But that does not mean it is not important!

There is someone on this board who poo-pooed my advocation of the use of a bitscope and possibly an RTA to examine the output of your DAW when you are working and mastering. His basic argument was: "If you can't hear it, then it's not important."

But the fact is that we mastering engineers work with unfamiliar material. Each day we get a new project. We very carefully listen to the source material and compare it to the result, but if the dither is incorrectly set it is not always intuitively obvious unless you turn it on, listen and then turn it off and say, "aha, there's a little loss in space and depth." To be honest, we don't have the time to make and prove by test listening on each and every job that we do something that we have already proved to ourselves time and again previously. So to that end, I recommend the visual aids, as protection against software bugs and human frailty.

How important is dithering? Around here I claim that I can hear a distinct difference in circumstances which you obviously are quite skeptical about. Frankly, that personal disagreement should not be important to either you or me!

If the theory (and in my opinion the practice) justifies good dithering practice, and people who you respect including leading psychoacoustic scientists feel that it is important, why should you be throwing a monkey wrench in the works to try to demonstrate how subtle (or inaudible at times) the phenomenon is? Of course it's subtle! We all know that. In many cases on the order of an audiophile difference, and in other cases there are people who claim to hear the difference in circumstances (such as the car) when it would be nearly impossible to pass a blind test even if their reactions are true.

I describe the sonic effects and when they are important and not important quite rationally in the second edition of my book.

Hope this helps,


BK

[twwalsh]

ethan - i checked out your test, and i have a few thoughts.

to make judgements about the effect of dither, we'd need to hear the original undithered 24 bit source. otherwise, we're comparing apples and oranges. also, since the arrangement changes, instruments come in and out, there's no way to know if a psychoacoustic change is due to an arrangement change or a dither change.

without knowing how many times dither was applied, there's no frame of reference or starting point. is it 20? 5? is the average length of time of a change 10 sec? 5 sec? 30 sec? i can hear edits in this material for sure...as well as other "psychoacoustic" changes but do you really expect me to spend a couple hours mapping out the results without any framework to work within?

i think a much more appropriate test would be to post a full 24 bit source file. then post the truncated full clip and the dithered full clip. have us choose which we think is dithered and which we think is truncated. AND which we prefer. this other test is a wild goose chase. it might be interesting but it certainly doesn't conclusively prove anything!

i'd like to add that my description of dither being a _blatant_ change should be taken within the context of me doing a lot of critical listening. i think it should go without saying that something an engineer (especially a mastering engineer) calls blatant may very well be _inaudible_ to the average listener! also, rock material with a lot of crashing cymbals and more competition across similar frequencies might benefit more from dither. this test is classical guitar with MIDI drums and instruments!

[Ethan Winer]

Originally Posted by twwalsh

to make judgements about the effect of dither, we'd need to hear the original undithered 24 bit source.

I promise you it is indistinguishable from my test file. The main reason I'm reluctant to post that is people could then null files and see where I turned the dither on and off. If my test is to be at all scientific, I need to ensure that nobody can cheat.

without knowing how many times dither was applied, there's no frame of reference or starting point. is it 20? 5? is the average length of time of a change 10 sec? 5 sec? 30 sec? i can hear edits in this material for sure...as well as other "psychoacoustic" changes but do you really expect me to spend a couple hours mapping out the results without any framework to work within?

If dither really mattered as much as "conventional wisdom" says it matters, it should be easy to pick out where the dither is used and where it's not. Maybe I should have been clearer in this thread, but my web page linked above says that I switched right in the middle of one musical phrase, so that alone should make it easy to pick out. I also said all of the switches are on beats or pickups. Also, no section is shorter than 7 seconds, but I realize now I said that only in Lynn's forum, not here or on my Dither web page. So I just added that to my web page.

this other test is a wild goose chase. it might be interesting but it certainly doesn't conclusively prove anything!

I disagree strongly. What it proves is that anyone who says dither is critically important, which is just about every magazine writer and forum poster, is full of crap. I have proven it is not critically important. I believe I have also proven it's not audible at all even under the most favorable conditions. But tell you what:

I will be glad to redo this test in any way you suggest and post a new file, except such that people can use trickery to see which is which. If I post a source 24 bit file, plus dithered and undithered versions, anybody can truncate the 24 bit file and know which file was which. If you can think of a way to post files that avoids all possible cheating, I'll be glad to do it. Not because I want to be proven right, but because I truly want to know what's right. Since you (and almost everyone else) says dither is audible, I welcome any appropriate test proposal that you design.

this test is classical guitar with MIDI drums and instruments!

Only the drums are MIDI. But why should that matter? Good MIDI drums are recordings of real drums. It's not like I used a cymbal patch on a MiniMoog!

--Ethan

[Ethan Winer]

Two more thoughts:

* I'll be glad to do this test using a source file you or anyone else provides, as long as it's not recorded at -40 dB etc.

* Another possibility is for me to test others here at my home studio and in my living room (I have two different excellent systems). That way I can play any files people want in any order, with the listeners not able to know which is which. Of course, this works only for those near to me.

--Ethan

[Ethan Winer]

Bob,

Originally Posted by bob katz

To be honest, we don't have the time to make and prove by test listening on each and every job that we do something that we have already proved to ourselves time and again previously. So to that end, I recommend the visual aids, as protection against software bugs and human frailty.

Understand that I am not opposed to the use of dither! Unlike unnecessarily high sample rates that reduce track counts and increase CPU load, dither is free. When this came up about jitter in Bruno Putzeys' forum at REP he did not dispute that jitter is inaudible, but he said he aims for the least amount possible anyway. And I agree with that. It's simply good engineering practice.

why should you be throwing a monkey wrench in the works to try to demonstrate how subtle (or inaudible at times) the phenomenon is?

That's a great question Bob. My sole motive is to learn the truth, and for others to learn the truth too. As you know, people argue all the time about what matters in audio, whether expensive speaker cables, expensive low-jitter outboard word clocks, expensive power "conditioner" products, and so forth. The key word is expensive. I see this mainly as a consumerist issue. If nobody can hear the difference between a $100 M-Audio card and a $3,000 high-end A/D/A, I believe consumers should know that.

From my perspective, things like image width and low end fullness have nothing to do with jitter or dither, but are acoustic in nature as explained HERE. As long as people know which they are listening to, it's impossible to separate fact from opinion. So all I'm trying to do is settle this stuff for once and for all. If it turns out jitter and dither really are audible in usual amounts, I want to know that too! Hence my quest for the truth via tests like this.

If I have any underlying motive, it's out of my frustration from seeing so many misguided magazine writers write about the importance of stuff like jitter, dither, and ultra-high bit depths and sample rates, while ignoring room acoustics that really does matter - at least 100 times more. Not because I want to sell more products! But because I believe it's the truth.

--Ethan

[twwalsh]

I'm not even remotely arguing that its critically important. Not sure where you got that idea. I was refuting your "proof" that it didnt matter at all. Your test is too open ended if you expect anyone to approach it realistically. I will write more when I'm not on my iPhone.

[jamsmith]

Originally Posted by Ethan Winer

As you know, people argue all the time about what matters in audio, whether expensive speaker cables

Do they still do this? I think it was Mix Magazine or EQ a dozen years ago gathered some "Golden Ears" for blind test on different speaker cables. The result was that none of the could really hear any difference.

Originally Posted by Ethan Winer

If nobody can hear the difference between a $100 M-Audio card and a $3,000 high-end A/D/A, I believe consumers should know that.

You make a real valid point. We really should be blowing the lid off on these. I myself and not nor will be rich. I for one need to know the most professional set up I can build for the least amount of money. I have been told that Cool Edit could never be used professionally. I will accept that that is true today. But when if first came out, compared to what was available for Windows at that time (short of maybe those $25,000 systems.), it stacked up and I saved a pile of money.

[jayfrigo]

Does dither matter? Absolutely. It is unassailable and provable scientific fact. You can measure it, the differences lie well within human auditory perception, and listening tests support it.

OK, that said, is it the most important thing facing us. By far, no.

Is it important in all situations? Again, no. Obvious all the time? For a third time, no.

If you are dithering from 48 to 24, you can still measure it, but you can't hear it. But why would you not do it? There's no downside, only upside. And who knows what will happen after repeated truncation, or with many tracks added together. It's just not worth omitting the dither.

Also, a lot of modern program material is so loud and so distorted that the relatively minor truncation distortion you add from not dithering is a tear in the ocean and likely will never be heard. The low level detail is where it really counts. Fades, reverb tails, the sound of a hall in an acoustic recording - these are places where it's most important, and especially taking these things to 16 bit.

Still, some perceive a slight brightening of the program due to truncation distortion, and occasionally some even prefer the extra distortion, though if I add distortion, I'd rather do it in a controlled manner somewhere earlier in tha chain. Then again, some perceive a brightening when they use some noise shaped dither as opposed to flat TPDF.

In the end, if you choose not to use it, that's your business. I choose to break the rules all the time if it suits the music (tubes, transformers, all kinds of fun distortion, and other creative techniques); but the key is to make an informed decision, understand what you are gaining and what you are losing, and not simply make an excuse for bad engineering practice.

[Ben B]

Originally Posted by jayfrigo

But why would you not do it? There's no downside, only upside. And who knows what will happen after repeated truncation, or with many tracks added together. It's just not worth omitting the dither.

Since this seems to be coming up repeatedly in this thread, I thought I'd point out that no one on any side of this discussion is advocating "not using" dither, at least the way I'm reading it. The only disagreements here are in terms of its importance to perceived audio quality in real-world listening conditions.

Ben B

[PettyCash]

Someone needs to figure out a way to do another fair test, by far this has been a very interesting subject.

I agree with the fact that most average consumers would not really be able to tell the difference or wouldnt even know what to listen for, but maybe some audiophile listeners out there would.

I think dither is important for me, because it makes sure that I have all my bases covered, and I just simply dont like to cut corners when it comes to working with audio.

If the next test proves that its not as critically important as people make it out to be, I would still use dither no matter what just for the security of doing so.

[Ethan Winer]

Originally Posted by twwalsh

I'm not even remotely arguing that its critically important. Not sure where you got that idea.

Here's where got the idea, plus many similar posts over at Lynn's forum:

Originally Posted by mpdonahue

I'll never forget the eye-opening experience of listening to F1 recordings we made as back-ups to the 1610 that sounded MUCH better than the master. The difference was Dither.

Originally Posted by Nishmaster

When truncating instead of dithering, you are effectively bashing those samples to digital zero ... This is that "edge" you hear, and it is completely audible and measurable.

Originally Posted by twwalsh

i can hear the 16 bit dither in voxengo elephant IMMEDIATELY and unmistakably

[Ethan Winer]

Originally Posted by jamsmith

Do they still do this? I think it was Mix Magazine or EQ a dozen years ago gathered some "Golden Ears" for blind test on different speaker cables. The result was that none of the could really hear any difference.

Belief in wires affecting sound quality is alive and well. Mostly among audiophiles, but I've seen it right here at GS in this Mastering forum.

--Ethan

[Ethan Winer]

Originally Posted by jayfrigo

Does dither matter? Absolutely. It is unassailable and provable scientific fact. You can measure it, the differences lie well within human auditory perception, and listening tests support it.

Of course it can be measured. It's the "listening tests" aspect I question Jay. This is why I'm asking for someone to supply a music fragment I can process and upload, since people don't seem happy with my music sample. If dither matters even a little, surely we can devise a test for folks to download, no?

And happy holidays to all! I REALLY don't like sounding argumentative on Xmas!

--Ethan

[jayfrigo]

Originally Posted by Ethan Winer

Of course it can be measured. It's the "listening tests" aspect I question Jay. This is why I'm asking for someone to supply a music fragment I can process and upload, since people don't seem happy with my music sample. If dither matters even a little, surely we can devise a test for folks to download, no?

And happy holidays to all! I REALLY don't like sounding argumentative on Xmas!

--Ethan

Should be able to test for it yes, but the Xmas thing does put a kink in it... Besides, I think we're all in agreement that it's pretty minor in the grand scheme. It's just that many minor things add up after a while, and just because it doesn't seem to matter on some program material, or on other monitors, doesn't mean we should hope that nobody ever listens to better music on better monitors. Besides, if we really wanted to do it right... nah, I don't want to open that can of worms, certainly not for a boring issue like dither.

If we're talking about MP3s on earbuds of modern, crushed music, then there's a whole bunch we need to fix first, before ever worrying about dither. Still, it is easy enough to do, and worth doing. It's funny that people will argue forever about higher sample rates capturing things that there is much scientific evidence that we can't hear, and then not wanting to use dither, which evidence shows we can hear.

Under similar conditions, many people here probably could not pick out the difference between an SSL and Mackie preamp, or know if an Apogee or Digidesign converter was used on a track. But they sure have definite opinons about them.

We work with subtle differences all the time, and consider them to matter. Why should this one be any different? I wouldn't want to over-stress the issue of dither, but one can't simply dismiss it. It's too well known and unerstood to be ignored.

And I don't think you're being argumentative. Feel free to continue the discussion. Happy holidays!

[bob katz]

Originally Posted by Ethan Winer

I promise you it is indistinguishable from my test file. The main reason I'm reluctant to post that is people could then null files and see where I turned the dither on and off. If my test is to be at all scientific, I need to ensure that nobody can cheat.

Dear Ethan: You are making some very wrong assumptions about how the double-blind test that you propose needs to be carried out. Au contraire: Your test is extremely UNSCIENTIFIC when judged by psychoacousticians and experts in double-blind testing. It violates all the rules of how to produce a sensitive double-blind listening tests.

For example, when the switching is unknown to the listener, it produces EXTREMELY skewed and unreproduceable results. In ALL valid double blind tests, the listener is told that a switch has occurred and the switch may be to the original material, or to the altered material, but the listener is NOT asked to try to guess where in the material the splice has occurred! This is entirely possible ONLY for EXTREMELY obvious material, not for subtle differences such as dither, which require an EXTREMELY sensitive test modality.

The test should be constructed to ensure success, not failure! The test you have constructed is destined to come up negative, even for the most sensitive of listeners on their own system in perfect acoustics!

1) Please read ITU-R Recommendation BS.1116, it outlines the procedures required for a sensitive double blind test

2) As per BS.1116, the test you will need to construct will have to consist of short segments of identical musical material repeated for the listener. In some types of tests, the first play is the "reference" and the second play is either a duplicate of the reference or the processed version. The listener is asked whether the sound is identical or different. As you can see, this is a fair test modality and can be scored quite accurately.

3) TRAINING is required. You have eliminated listener TRAINING, which is absolutely necessary for such a test to succeed. The short original sample is played (e.g the 24 bit original), and the result (either truncated or dithered to 16). The listener is allowed to spend time training his ears. This is not deceit, this is a REQUIREMENT for any test of sensitive material such as you propose.

4) The musical material has to be carefully chosen by listeners who claim to know the sound character that they are trying to identify (e.g. dithered vs. undithered), not be randomly chosen by the skeptic (you)!

And there are far more requirements which once you study them should become straightforward. Such a test is EXTREMELY expensive and time-consuming to construct and generally should be conducted by a trained psychoacoustician who needs to make sure that all factors are done properly.

I can almost guarantee that if you conduct this test in this manner, properly done, that you will find that dither does matter. I'd be happy to invite you to my studio any time to take (or give) a test in this manner.


Best wishes,


Bob

[bob katz]

Dear Ethan:

I also suggest you find the article: "Subjective Evalation of Large and Small Impairments in Audio Codecs", by Gilbert A. Soulodre and Michel C. Lavoie. From the AES 17th International Conference on High Quality Audio Coding. The authors compare the BS.1116 method versus other sensitive methods and more relevant to your proposal---- one of the tests they conducted involved having subjects judge "audio impairment" when calibrated amounts of noise was added to material.

They then measured statistically with error bars the subject's results and came up with a test methodology that is extremely sensitive.

In the end, if you are trying to judge "how subtle dithering is", you have to perform the test correctly!

BK

[bob katz]

Originally Posted by jayfrigo

Under similar conditions, many people here probably could not pick out the difference between an SSL and Mackie preamp, or know if an Apogee or Digidesign converter was used on a track. But they sure have definite opinons about them.

snip

And I don't think you're being argumentative. Feel free to continue the discussion. Happy holidays!

I agree Ethan is not being argumentative, but he is being misguided in his attempt to make a test where the listener is at a complete disadvantage. I could produce a test, too, with random splices between +3 dB at 5 kHz EQ versus flat, and if the loudnesses of the two were matched, it would be EXTREMELY difficult to hear it, because MUSIC changes, it is not constant, and the listener has not been given a comparison between the SAME material repeated two different times in two different ways!

If,however, the source were pink noise, then EVERYONE could hear the difference in the two EQs. I'm not arguing that dither is not subtle, it is, very subtle, but I am feeling that if Ethan were to make a trip to a good mastering studio with good acoustics, one where we exercise these decisions ever day, even he would be shocked at how much more audible the differences are than he currently believes. But it is subtle :-), a lot more subtle than the claims being made on Lynn's forum that Ethan has quoted, so to that extent I agree with Ethan. He just has retorted with a proposed test methodology that is completely invalid!

And since this subtle difference will be extremely difficult to quantify or prove scientifically (been there/done that/already discussed it with one of the most prominent and respected psychoacoustician and dither designers (Bob Stuart