The No.1 Website for Pro Audio
Good dither practices, what are yours?
Old 5 days ago
  #1381
Gear Maniac
 

Quote:
Originally Posted by J3XS View Post
Quote:
Originally Posted by FabienTDR View Post
I guess you got it, truncation DOES NOT accumulate dangerously. In fact, at extremes, the opposite happens! The more numerous the truncations, the more likely the chance that they will self dither (in sum, as a system, with many many many little truncations. Only reason we don't see this beneficial effect here is because we're using a pure sine).
Dang, that sounded like me, except for the bold and exclamation mark, which I avoid in forum discussions. I had to double check it wasn't me. (And I guess I would have said they tend toward white noise, instead of "self dither", especially because the latter sets people off, but they'd sound the same at limits and that's what he means.) Way to go Fabien.
Old 5 days ago
  #1382
Gear Nut
 

Quote:
Originally Posted by earlevel View Post
Dang, that sounded like me, except for the bold and exclamation mark, which I avoid in forum discussions. I had to double check it wasn't me. (And I guess I would have said they tend toward white noise, instead of "self dither", especially because the latter sets people off, but they'd sound the same at limits and that's what he means.) Way to go Fabien.
But it's still all theoretical as there is still a conflict of interest from equally qualified audio engineers!

e.g. why stop at a sine wave and then speak theoretically? Can this not be done with actual recordings?

Where are the songs where these differences were heard / proved? Can they not be shared?
Old 5 days ago
  #1383
Lives for gear
 

Quote:
Originally Posted by J3XS View Post
But it's still all theoretical as there is still a conflict of interest from equally qualified audio engineers!

e.g. why stop at a sine wave and then speak theoretically? Can this not be done with actual recordings?

Where are the songs where these differences were heard / proved? Can they not be shared?
You beat me to it.

I was thinking this weekend of preparing as series if clips mixed through my analogue chain out of a HEDD 192 with and without 24bit dither.

People can then download them if they wish, process them at will with EQ etc to reveal, for example, the all telling digital crunchiness truncation noise can cause with subsequent processing and then ....

Name which files have 24 bit dither.

It's not the dreaded ABX test, just a critical listening experiment.

It'll be interesting, unless of course the files can be digitally analysed to reveal which have 24 bit dither and which don't?

In which case it's a waste of time.
Old 5 days ago
  #1384
Airwindows
 
chrisj's Avatar
Quote:
Originally Posted by J3XS View Post
I'm just wondering what about post processing from e.g. streaming services (i.e. adjusting the LUFS levels themselves which is very common)?

Also stuff like iTunes adding their own EQ types (hip-hop, electronic settings)?

Even if the end song is dithered, will it be truncated again for their own processing? (since it's not a physical medium like CD > DA)
Never said modern digital wasn't garbage

Note that people do not PAY for modern music! You may say they do. They end up paying the computer guys, digital service holders. They ain't paying the artists. Nor are they getting a product that particularly connects 'em with an artist. It's wallpaper.

As for multiple truncation buildup turning into self-dithering, no, of course not.
Old 5 days ago
  #1385
Lives for gear
 

Quote:
Originally Posted by chrisj View Post
Nor are they getting a product that particularly connects 'em with an artist. It's wallpaper.
I'm sorry, with the greatest of respect, that's a bit of a gross generalization based on a highly subjective view.

My wife and kids listen to their favourite artists and music on Spotify (a streaming service) and I can assure you they are very much connected to the artists they listen to - for them their music is definitely not wallpaper!!

This thread had been an "audiophile" technical debate amongst a tiny pier group that have a deep passion about audio, of which I don't doubt your considerable knowledge and expertise.

Its not a thread, imho, even remotely related to "the entertainment business" and the consumption of popular music.

Last edited by thehightenor; 5 days ago at 01:24 PM..
Old 5 days ago
  #1386
Lives for gear
 

Quote:
Originally Posted by thehightenor View Post
You beat me to it.

I was thinking this weekend of preparing as series if clips mixed through my analogue chain out of a HEDD 192 with and without 24bit dither.

People can then download them if they wish, process them at will with EQ etc to reveal, for example, the all telling digital crunchiness truncation noise can cause with subsequent processing and then ....

Name which files have 24 bit dither.

It's not the dreaded ABX test, just a critical listening experiment.

It'll be interesting, unless of course the files can be digitally analysed to reveal which have 24 bit dither and which don't?

In which case it's a waste of time.
Can I suggest using the best possible high resolution "direct to digital" classical recording you can source, featuring music with lots of quiet periods?
Old 5 days ago
  #1387
Lives for gear
 

Quote:
Originally Posted by diggo View Post
Can I suggest using the best possible high resolution "direct to digital" classical recording you can source, featuring music with lots of quiet periods?
I'm on this thread speaking as a producer of "popular music" with bass, drums, guitar, piano and the normal rock "n" roll instruments, so it makes sense for me to be focusing on the actual programme material I would be applying dither to.

24 bit dither does matter in rock 'n' roll doesn't it?
Old 5 days ago
  #1388
Lives for gear
 

Quote:
Originally Posted by thehightenor View Post
I'm on this thread speaking as a producer of "popular music" with bass, drums, guitar, piano and the normal rock "n" roll instruments, so it makes sense for me to be focusing on the actual programme material I would be applying dither to.

24 bit dither does matter in rock 'n' roll doesn't it?
Of course, but if you want the best chance of deriving an audible result, use music which is very susceptible to degradation which is audible.

Remember, not all of us only work with rock n roll - my mastering workload varies from extreme Post Rock to string quartets.

But OK, include ride cymbals, high quality acoustic guitars and plenty of light and shade dynamics. That's one thing i always have to deal with when mastering the various dynamic extremes with Post Rock.

Maybe throw in a few shakes of your car keys just for jollies.

Set the test up to succeed, not to fail.
Old 5 days ago
  #1389
Lives for gear
 

Quote:
Originally Posted by diggo View Post
Of course, but if you want the best chance of deriving an audible result, use music which is very susceptible to degradation which is audible.

Remember, not all of us only work with rock n roll - my mastering workload varies from extreme Post Rock to string quartets.

But OK, include ride cymbals, high quality acoustic guitars and plenty of light and shade dynamics. That's one thing i always have to deal with when mastering the various dynamic extremes with Post Rock.

Maybe throw in a few shakes of your car keys just for jollies.

Set the test up to succeed, not to fail.
I didn't realise genre could have an effect on the audible result!

I got the distinct impression from this thread, dither and it's positive effects are genre neutral.

I'm not hoping for a positive or negative result, I just thought is would be interesting to move the thread into the arena of observable practical science.

That said, I'm waiting on a reply as to whether an audio file can be digitally analysed to reveal if it contains 24 bit dither, does dither leave a digital finger print?

For if it does, then with human nature being what it is - there's no point in bothering posting any files.
Old 5 days ago
  #1390
Lives for gear
 
ben_allison's Avatar
I use Logic and stay mostly in the box, is there a TL;DR for the thread?
Old 5 days ago
  #1391
Gear Maniac
 

Quote:
Originally Posted by J3XS View Post
But it's still all theoretical as there is still a conflict of interest from equally qualified audio engineers!
Of course it is. But there are good reasons it's true—it may be theory that thermal noise is white and gaussian and at absolute minimum ~131 dBu (one resistor!) for audio gear at room temperature. But show me anything that doesn't have it. If you could, that would be a black swan that would tip the physics world on end. Yet...it's just theory.

Anyway, I'm not getting into that again, was just noting I'm not the only one that says this. (Also, Fabien and I disagreed on what samples represent, was nice to see we have the same view on this point.)
Old 5 days ago
  #1392
Gear Maniac
 

Quote:
Originally Posted by ben_allison View Post
I use Logic and stay mostly in the box, is there a TL;DR for the thread?
If you stay in the box, it's simple:

1) For 16-bit, dither

2) For 24-bit, not possible to hear it

3) But for 24-bit do it anyway—if you're ITB, the only time you need to drop to 24-bit is typically for the final, just like 16-bit, it's the last step for distribution—might as well do it as your same workflow step as for 16. If you forget this step for 24-bit, don't fret, carry on. No way to tell the difference by listening anyway.

IMPORTANT: If #2 bothers you, copy the above text, delete #2 , renumber step 3 to 2. Read it. If you want a shorter list without consequences, delete 2 and 3.
Old 5 days ago
  #1393
Lives for gear
 
bogosort's Avatar
Quote:
Originally Posted by thehightenor View Post
That said, I'm waiting on a reply as to whether an audio file can be digitally analysed to reveal if it contains 24 bit dither, does dither leave a digital finger print?
No, for recorded audio there's no litmus test for 24-bit dither.

It's trivial to see whether dither has been applied to lab-type signals like digitally-generated sine waves, as the quantization noise is fully correlated with the signal. If you look at just the LSB of a quantized and undithered sine wave, you'd see something like a pulse-density modulated stream of 1s and 0s, with alls 1s in the region of the sine's positive peak, all 0s near the sine's negative peak, and rapidly alternating 1s and 0s near zero-crossings. Dither kills this correlation, leaving you with a bunch of random 1s and 0s.

Now consider the same experiment with 24-bit recorded audio. The LSB measures deep into the analog noise floor, so the stream of LSBs will look random regardless of whether dither has been applied (which is precisely why dithering a 24-bit analog send is pointless).

I don't think a statistical analysis would help. If we treat the LSB stream as a random variable sampled from a probability distribution, I'd expect both the dithered and non-dithered LSB to look roughly Gaussian.
Old 4 days ago
  #1394
Lives for gear
 

Quote:
Originally Posted by bogosort View Post
No, for recorded audio there's no litmus test for 24-bit dither.

It's trivial to see whether dither has been applied to lab-type signals like digitally-generated sine waves, as the quantization noise is fully correlated with the signal. If you look at just the LSB of a quantized and undithered sine wave, you'd see something like a pulse-density modulated stream of 1s and 0s, with alls 1s in the region of the sine's positive peak, all 0s near the sine's negative peak, and rapidly alternating 1s and 0s near zero-crossings. Dither kills this correlation, leaving you with a bunch of random 1s and 0s.

Now consider the same experiment with 24-bit recorded audio. The LSB measures deep into the analog noise floor, so the stream of LSBs will look random regardless of whether dither has been applied (which is precisely why dithering a 24-bit analog send is pointless).

I don't think a statistical analysis would help. If we treat the LSB stream as a random variable sampled from a probability distribution, I'd expect both the dithered and non-dithered LSB to look roughly Gaussian.
Thank you for that detailed reply and so clearly explained.

I have not only enjoyed this thread but genuinely learnt some great information and explanations about aspects of digital audio I've not considered before.
Old 4 days ago
  #1395
IMHO quantization "damage" is best explored by listening to the difference between a truncated version and its original.

This signal can be amplified, listened to, it can be analyzed, without being distracted by your favorite lyrics, chords or a funky drum solo. It fully represents the error (i.e. what you hear is exactly what is being "added" to the audio).

If no regular patterns, no traces of or similarities with the original can be found in the difference signal, then it is likely fully de-correlated, and very likely purely consists of white noise. Quantized white noise is white noise, meaning there's no need for an additional dither.

You can double check the lack of patterns easily with a spectrum analyzer, but in fact, the ear is an excellent corelation analyzer, one of the few things it does really well.

It leads to a rather surprising lesson (or at least, it did for me). I found it rather difficult to find any bad examples. The true worst cases appeared with "fades to zero", other types of "modulation to zero", or with simple, digitally generated signals (= arguably not music). But my best sounding records easily truncated down to 12 or 8bit, without ever producing measurable or audible distortion. Worth trying out. Reaper has a free truncation plugin, it's great for quick tests.

In standard truncation cases, i.e. simple format conversion, the truncation will often happen deep below the material's natural noise floor. And this on top of the fact that the truncation of a wildly moving wideband signal by itself can be seen as a crude form of random number generator.

The only reason why all this isn't worth the headaches is that dither noise really never hurt anybody (given it's not a fancy trademarked type of dither, but one that accumulates gracefully when added several times).

Last edited by FabienTDR; 4 days ago at 01:08 AM..
Old 4 days ago
  #1396
Gear Maniac
 

Quote:
Originally Posted by FabienTDR View Post
IMHO quantization "damage" is best explored by listening to the difference between a truncated version and its original.

This signal can be amplified, listened to, it can be analyzed, without being distracted by your favorite lyrics, chords or a funky drum solo. It fully represents the error (i.e. what you hear is exactly what is being "added" to the audio).

If no regular patterns, no traces of or similarities with the original can be found in the difference signal, then it is likely fully de-correlated, and very likely purely consists of white noise. Quantized white noise is white noise, meaning there's no need for an additional dither.

You can double check the lack of patterns easily with a spectrum analyzer, but in fact, the ear is an excellent corelation analyzer, one of the few things it does really well.

It leads to a rather surprising lesson (or at least, it did for me). I found it rather difficult to find any bad examples. The true worst cases appeared with "fades to zero", other types of "modulation to zero", or with simple, digitally generated signals (= arguably not music). But my best sounding records easily truncated down to 12 or 8bit, without ever producing measurable or audible distortion. Worth trying out. Reaper has a free truncation plugin, it's great for quick tests.

In standard truncation cases, i.e. simple format conversion, the noise floor will often be skyscrapers beyond the material's natural noise floor. And this on top of the fact that the truncation of a wildly moving wideband signal by itself can be seen as a crude form of random number generator.

The only reason why all this isn't worth the headaches is that dither noise really never hurt anybody (given it's not a fancy trademarked type of dither, but one that accumulates gracefully when added several times).
That pretty much echoes my sentiments—particularly the being surprised how it wasn't as easy to produce noticeable notable distortion from actual music as expected. (Though my experience is that it gets a lot harder starting with about 14-bits, at least in things like fades—you can see and hear this is my video while analyzing an arbitrary snip of music.)

You also seem to have basically the same feeling I have, that if—for a particular recording—the isolated truncation error sounds like unmodulated white noise (that is, to the ear the same as it the residual from dithering), then dithering doesn't improve anything. But be aware, I've found that audio folks are pretty, um, suspicious in this regard, and I've had mastering engineers literally theorize that truncation is so special that the brain can subconsciously pick it out as causing an unnatural sound, even far below the signal and the noise floor. This isn't logical, of course—if it sounds like broadband uncorrelated noise to the ear, adding it to a pristine signal (and therefore the equivalent of truncating) is not going to change that. If audio behaved in such a bizarre way, mixing audio would be pretty troublesome instead of a simple summation.
Old 4 days ago
  #1397
Airwindows
 
chrisj's Avatar
The catch there is, the 'difference' audio can resemble white noise all you like but it will be trivial to ABX an objectionable quality in the 12 bit or 8 bit versions of your truncated file.

I would say the explanation is, if you subtract the literal correlation source from the correlated noise, you're removing the context that makes the truncation irritating. If it's welded to a very complicated sound, the ear can treat it like some kind of mechanical 'buzz' attached to the audio, but if you remove the sound and leave only the buzz, it's too hard to synthesize what's 'supposed' to be there and you lose the connection that's otherwise obvious.

Maybe there's a creative use for this phenomenon
Old 4 days ago
  #1398
Lives for gear
I can't believe there are so many posts on this subject. That's why I am adding one more.
Old 4 days ago
  #1399
Gear Maniac
 

Quote:
Originally Posted by chrisj View Post
The catch there is, the 'difference' audio can resemble white noise all you like but it will be trivial to ABX an objectionable quality in the 12 bit or 8 bit versions of your truncated file.

I would say the explanation is, if you subtract the literal correlation source from the correlated noise, you're removing the context that makes the truncation irritating. If it's welded to a very complicated sound, the ear can treat it like some kind of mechanical 'buzz' attached to the audio, but if you remove the sound and leave only the buzz, it's too hard to synthesize what's 'supposed' to be there and you lose the connection that's otherwise obvious.

Maybe there's a creative use for this phenomenon
OK, but you're doing what I said already warned Fabian about ("But be aware...I've had mastering engineers literally theorize that truncation is so special that the brain..."). Not only is there no proof of this, but we already know that audio signals in general sum just fine. There are well-known and studied phenomena, such as masking, but if this "context of noise" thing existed, you could probably point to research on it.

Not only that, but the really fun stuff we do these days with source separation would likely not work so well, as the context is always lost.

But OK, I get that you might counter that the noise is part of the audio at that point, so it's different. But let's think about that. We start with the audio we'd like to present, we truncate it. We've introduced error—what we know for sure is the error bounds, with is [0,-1) lsb with simple bit truncation.

We take the difference between the truncated and the original, that's the isolated error. For 24-bits, we won't be able to hear the error, but since we've isolated it we can boost the gain as much as we want. Let's say we listen to the error and it happens to sound like smooth white noise all the way through. If so, the correlation with the music must be poor. (We'd like the error to be indiscernible from the same test dithered. What Fabien and I have said is that we found this to be true with higher bit depths with real recordings. Just saying that, from experience, I don't think I'm talking about an unlikely situation.)

Why would you expect that summing this error back with the original signal (bitwise identical with the truncated version) makes us no longer hear the simple sum of original plus smooth white noise?

Again, I think if this were a real phenomenon, you could point to research. Or at least devise a hearable test case. Otherwise it seems a little like when people swear that if they eat this one ounce dessert they'll gain five pounds
Old 4 days ago
  #1400
Lives for gear
Quote:
Originally Posted by diggo View Post
Can I suggest using the best possible high resolution "direct to digital" classical recording you can source, featuring music with lots of quiet periods?
Which would make it easy to spot the dither as a kind of low level "tape hiss".
Old 4 days ago
  #1401
Lives for gear
Quote:
Originally Posted by thehightenor View Post

It'll be interesting, unless of course the files can be digitally analysed to reveal which have 24 bit dither and which don't?

In which case it's a waste of time.
As long as you dither the final mix and there aren't any silent or very quiet passages in the material I can't think of a way of analysing the audio to spot dither (on the individual tracks that is).

Last edited by Scragend; 4 days ago at 09:12 PM.. Reason: clarification.
Old 3 days ago
  #1402
Motown legend
 
Bob Olhsson's Avatar
 

Properly implemented dither should be less audible than truncation distortion. Masking is a moving target that depends, among other things, on the listener's hearing damage. The more damage, the less masking happens.
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