Dither - My Ears can't hear it

Status
Not open for further replies.
In other news, today I scored another box of rigid fiberglass. I have already deployed 3 boxes in the room (2 of them 4"!), but while the mix position is just ducky, the back of the room gets a little squirrelly.

I was out of space to hang from walls and ceilings, so I'm building a portable 3 panel Z-fold gobo :) Again, each panel 4" thick.

That ought to be just the ticket!
 
In other news, today I scored another box of rigid fiberglass. I have already deployed 3 boxes in the room (2 of them 4"!), but while the mix position is just ducky, the back of the room gets a little squirrelly.

I was out of space to hang from walls and ceilings, so I'm building a portable 3 panel Z-fold gobo :) Again, each panel 4" thick.

That ought to be just the ticket!


Are you using Roxul ?

Or is that different ?
 
Or come up with your own better tests if you think mine are flawed.
Your tests are well-intentioned but very colored and extremely limited. All you have demonstrated there that one person reports that he could not hear one type of dither (you never mention which one, BTW) on one particular song on one particular system. I'll buy that. I never claimed otherwise. I'm not weeping over that result any more than I am weeping over not seeing Ms' Delaney this weekend, because I know that neither one can be taken very seriously.

I have already defined here many of the major parameters required for a true test of a subject so marginal and so controversial. I also have offered to design, build and host such a test, and accept it's results however they come out. I have also said that it will just have to wait for a little while because I have other work to get out of the way first.


G.
 
I'm bumping this because I have new information.

I see nothing new here Ethan. I think Lynn summed up this issue nicely:

"I would like to add that no one here should be debating the audibility of dither. It's clearly audible, if you've ever done any tests reducing shorter word lengths. I think the discussion here is focusing rather on the threshold of audibility of dither.".

Sometimes you can hear the difference, sometimes you can't. That's pretty much the end it. If the effects are masked with other types of harmonic distortion and noise, or you lower levels, it's much more difficult to hear, but it's still there. Just because one sprays the bathroom with deodorizer it doesn't mean that there isn't crap in the air.
 
More info regarding this thread:

http://www.gearslutz.com/board/mastering-forum/163362-dithering-101-a.html

We could debate this ad-nauseum, as for me I'm done.
Interesting quote from Ethan himself in that thread:
Ethan said:
Absolutely! I'm not arguing against dither. I'm merely arguing that it's not as important as most people believe it is.
Ah, now this is a different story. It's one thing to say that dither is meaningless. It's another to say that it's not meaningless, just not a huge earthshaker. I'm not sure too many folks would disagree with that second one, actually.

What the argument boils down to is the dichotomy between two perspectives, both presented very sloppily:

On the one side is the view that dither is irrelevant because it is, at best, at the very threshold of audibility. They defend that viewpoint with very poorly designed listening tests that actually tell us very little other than in one very narrow, very specific instance, audibility is in question.

On the other side are those that believe that if something causes a measurable difference, it's relevant, even if it's audibility is, at best, subtle or intermittant. And since it can make any difference whatsoever, it would be bordering on negligent not to do it. They back their arguments with pictures and digital theory that actually tell us very little other than dither is a measureable effect, but says nothing about it's actual audiblity. There are a few who claim to have platinum ears who claim they can hear and ID dither anywhere and everywhere, but these are usually the guys that say they can hear a pin drop in a vacuum.

All those tests and polls on both sides tell us nothing definitive. Let's face it, if any of them did, the "argument" would be over.

The problem is, dither is not a definitive, black/white thing. It never will be. It will always be a game of percentages. This is why the argument will never stop. It is sometimes audible to some, sometimes not. Sometimes sensed by others, sometimes not. Some people wouldn't/couldn't hear it if it clubbed them over the heads (lets face it, we live in a society that for the most part can't hear or doesn't notice a difference between a 24-bit WAV and a 128k MP3.)

But as long as there is a core - no matter how large or small - that claim they can sense the difference, it seems to me that all the rest is irrelevant. There is, at that point, no reason for the detailers and technicians to discard the process, or to even argue about it.

G.
 
I have already defined here many of the major parameters required for a true test of a subject so marginal and so controversial. I also have offered to design, build and host such a test, and accept it's results however they come out. I have also said that it will just have to wait for a little while because I have other work to get out of the way first.

Excellent. When you're ready let me know, and I'll be glad to host the test files if you need that.

--Ethan
 
Ethan, I like your triangle test. I fell compelled to make a couple of observations:

First, I think it's a great example of why measurement and analysis can be arguably more important than listening. Everybody is quick to run to listening tests, and of course at the end of the day, listening is what matters. However, listening tests are often extremely poor at pinpointing the differences, whereas analysis is often quite good.

It's even easier with something like a single triangle hit; a simple FFT analysis I did before I finished reading the thread had me wondering about the Apogee hum. I hadn't listened to the files, but it was quite clear. And there it was being observed in the listening test.

I personally feel that a single triangle hit is probably a better measure of a converter than 5 minutes of program material, which is what one tends to see in often not too well designed tests. So I applaud that.

Second, I have to say it seems not all Soundblasters are created equal. I am actually a Soundblaster fan; even so, if I can find my triangle, I can show you that my SB--a garden variety 2005 vintage--doesn't do so well over 16kHz. And I have tested other converters or even SRC routines (which are just another flavor of the same filtering an A/D has to do) that perform better or worse with very high-frequency content.

As for dither--I am confident based upon my earlier analysis that I could prove audibility, but then I would have to subject you to my flute-playing, and you'd have to listen on headphones. I agree you probably couldn't detect it on a mix, but I don't think use of headphones is beyond the pale, even on solo flute music.

I think that's my main trouble with your argument, earlier you say " . . . normal listening levels . . ." when I think that would have to be "proper listening levels". But there are a lot of people, if not most, who listen on headphones well above proper levels, and that's where I think it will stand out.
 
Excellent. When you're ready let me know, and I'll be glad to host the test files if you need that.
Uh uh. Nobody gets access to the test files except via the testing methodology. Besides, I would want and expect you to participate in the test, Ethan. Those with access to any of the test materials whatsoever will be excluded from participation in the test itself.

Furthermore, the design of the test is such where I have to have programmatic access to the files; I could not do that with files on your server unless I had administrative access rights to your server, which I neither want nor ask for.

There remains several BIG problems with any Internet-based testing methodology:

There's a whole lot of people who won't give answers without running forensic testing on the files first. This would completly invalidate their results. Second, the internet is just chock full of vandals who love to fuck with these kinds of things just for the sake of fucking with them. Remember Sanjaya?

So I'm still thinking this through. I'd love to see a real test done at testing centers where the files themselves remain unatainable to both the tester and the testee, and is a process that its just too much of a hassle for your average Internet vandal. I don't know how to do that over the Internet, though.

Nevertheless, any such REAL survey testing, even over the net, has GOT to be better and more accurate than any 10 tests or polls put together of the type I have seen so far.

G.
 
How can you tell anything from that triangle test. The Apogee file sounds horrible. You are 100% right, based on that test, I'd go soundblaster all the way.....................

I'll let Ethan and Bob Katz fight it out. :D
 
listening tests are often extremely poor at pinpointing the differences, whereas analysis is often quite good.
Agreed. Ears and perception change quite a lot over the course of a few seconds. Test gear gives the same results every time (assuming you know how to use it.)

I personally feel that a single triangle hit is probably a better measure of a converter than 5 minutes of program materia
Indeed.

I have to say it seems not all Soundblasters are created equal.
Agreed there too. The noise floor of my current SoundBlaster is about 10 dB lower than the last one I had from 5+ years ago.

As for dither--I am confident based upon my earlier analysis that I could prove audibility, but then I would have to subject you to my flute-playing, and you'd have to listen on headphones.
No problem - go for it! :D

there are a lot of people, if not most, who listen on headphones well above proper levels, and that's where I think it will stand out.
I listen loud too, and requiring headphones is fine. If you tell me to play the file as loud as I can reasonably stand, I'm glad to do that. I'll cry foul only if I have to turn up the volume even louder than that on a fade-out or reverb tail to hear the use of dither. Or if you record your flute so it never gets above -30 or some such. If you believe you can demonstrate the value of dither at loud but tolerable levels, please post it. Not just for me, but for everyone else here too.

--Ethan
 
Nobody gets access to the test files except via the testing methodology.

That's fine with me. I agree that as soon as people can look at the files with an FFT, or subtract one from the other to expose just the dither, that means they can cheat. And if they can cheat, you can be sure some will cheat just to prove their point. This is a big problem with 'net tests.

That said, I'd be glad to see any pair of files that shows dither is audible at all. No blind test needed - just post some files that obviously sound different, where the only difference is one is dithered and the other is truncated.

--Ethan
 
again...why is it such a big deal? Most DAWS will apply dither automatically when bouncing to a lower bit rate (most editing software does that as well) so it takes an extra 2 seconds, even if it's a theoretical improvement that one person might be able to hear...why not? It's not like those 2 seconds could be spent doing anything else more important.
 
No problem - go for it! :D


I listen loud too, and requiring headphones is fine. If you tell me to play the file as loud as I can reasonably stand, I'm glad to do that. I'll cry foul only if I have to turn up the volume even louder than that on a fade-out or reverb tail to hear the use of dither. Or if you record your flute so it never gets above -30 or some such. If you believe you can demonstrate the value of dither at loud but tolerable levels, please post it. Not just for me, but for everyone else here too.

--Ethan


Well, after trying it, I'm going to have to concede. I just can't play a wind instrument quietly enough (while still having appropriate levels for the loudest parts) to manage it. I switched to recorder as I was doing a bit better there, but I only managed about a 15dB range. There is no way that is going to do the job for such a test.

So instead, I tried to replicate my earlier test using a sine wave + pink noise acoustically. And I have to say I did a pretty good job on the reproduction. Even so, I couldn't find the quantization distortion! It should be there, but I can't generate it! And I tried pretty hard!

Have a look, this is the test signal, 1kHz at about -52dBFS, with 20dBA equivalent pink noise (the file I posted earlier, which apparently you couldn't read), in yellow. The quantization distortion peaks at 3, 5, 7kHz are pretty clear, and they are audible on cans too.

But the blue is the tail of me banging a single note (C6) on my piano. As you can see, its level is quite lower, -80dBFS. And it is still easily audible. Also, you can see I matched the test noise floor rather closely (this file was totally unprocessed). In theory, that should produce quantization distortion (although the level may have dropped too much at -80dBFS).

The peaks in blue you see are hum, and are consistent irrespective of source signal level. There is no measurable harmonic distortion above the noise floor. The peak of the note attack was -6dBFS. The microphone was a KSM141 (on omni), a couple of feet from the spinet piano.

Thus I surmise that a sustain above -80dBFS must be required to measure quantization distortion. However, in this case, when the note had not faded to that level, but was around -60dBFS, its own overtones dominated any potential quant distortion. In fact, if you watch an FFT during the entire note decay, no differences can be detected between the original and truncated file. So not only would a sustained note at a somewhat higher level (but not too high!) be required, but it would also have to possess few overtones.

I have verified that proper truncation is occurring and no dither is applied on truncation by performing the same procedures on test signals, and measuring the resulting distortion.

This one is a mystery to me. I believe I should have been able to demonstrate some distortion, based on theory and practice with very similar test signals, but I cannot.
 

Attachments

  • dither test.GIF
    dither test.GIF
    36.7 KB · Views: 53
Here's a better graph, I redid the test signal at a closer level to a point in the piano fade, and selected a 1 sec segment of each for analysis. The test signal (again, including pink noise) produces clear quant. distortion . . . the piano does not :confused:
 

Attachments

  • better dither test.GIF
    better dither test.GIF
    33.5 KB · Views: 51
Thinking out loud . . .

New theory, the piano results, even with similar self-noise in the A-weighted range, actually are noisier because of 60Hz peaks (in the interest of space, I cropped that out of the picture). Thus there is a form of "noise-shaping" occurring with the piano signal . . . it just so happens that its noise is audible, in the lower frequencies :o


Yep, a quick retest of the test signal with added simulated 60Hz hum . . . no more quant. distortion.

This redistribution of wealth is trickier than I thought :(
 
Status
Not open for further replies.
Back
Top