Topic by kalessin

  • Avatar
  • kalessin
  • Member
  • Offline
  • From: Germany
  • Registered: 06-12-2013
  • Posts: 306

Topic: Bit depth, noise, dynamic range

Jake Johnson wrote:
vjau wrote:

That's the same thing.
Number of possible amplitudes = dynamic range.
If you don't agree think photography.

No, no. Range indicates the lower and upper limits. Bit depth indicates the number of amplitude levels possible at a given instant. Very different things, aren't they?

EDIT: In the way that I would use the terms, in terms of photography, a low resolution black and white digital image might have a range from white to black, but only have ten gray-scale steps between them. A higher resolution image might have the same range from white to black, but have thousands of gray-scale steps between them.

vjau is actually right. The number of possible different values that can be present in a signal is exactly what dynamic range means. You are thinking of something different. I'll try to explain below.

Jake Johnson wrote:

Your original argument seemed to be that the main effect of a higher bit rate is an increase in dynamic range. I was saying that the more significant audible contribution was instead simply the number of amplitude levels simultaneously available.

I'll assume you mean bit depth and not rate. And yes, I said that: higher bit resolution means lower quantisation noise means more dynamic range, and I stand by that statement.

Concerning images: the theoretical maximum dynamic range of a 8bit JPEG image is (*drum roll*) 8 stops. Yes, that's just 8; not 10, not 12. In fact it is closer to 6-7 in practise, and many printers are unable to even transfer that to paper. This value corresponds nicely to the human eye, which (to my knowledge) has a static contrast range of just about 6.5 stops. In other words, there's a reason we call 8bit RGB images 'true colour' images.

On the other hand, modern cameras claim to have 'dynamic ranges' of 12 stops and more. So, how does this work? The simplified answer is that camera manufacturers misuse the word 'dynamic range' because it sounds impressive. They actually mean the absolute brightness range the camera captures and compresses into the image. 'Compression' is the important point here: a photo is always the result of (quite heavy) dynamic range compression. This means that yes, of course you can capture absolute contrasts of 12 EV into a single image. But local low-contrast details will still vanish.

Were an image a linear representation of brightness values (i.e., uncompressed), then the lowest relative brightness we could encode in 8 bits would be about -8EV. Local details would be theoretically visible anywhere if they have a contrast that is higher than -8EV relative to 'white' (i.e., the maximum brightness). In numbers: a local feature has to have a contrast difference of at least 1/255 of the white value in order to be visible. This is disregarding colour information, which we can (and do) use to improve contrast perception. Since images in practise are quite heavily compressed in terms of dynamics, the minimum contrast of a small detail has to be even higher than -8EV. This corresponds to the simple fact that dynamics compression decreases the dynamic range of the data, not increases it. Everything is squashed together, details get washed out.

So, why do camera manufacturers compress dynamics? Because our eyes do it as well (but we are better at it). While the static contrast range of the human eye is about 6.5 stops, it can adjust to different brightness ranges rather quickly: a few stops in about a second or less (by closing/opening the iris), and more if allowed to adjust for longer times. This is the dynamic contrast range of the eye, which is a lot higher than 6.5 stops. Which effectively means that we permanently look at different parts of the landscape, our eyes adjust accordingly, and our brain puts it all together. Because of this, images with about 10-12 stops of compressed brightness range actually look more natural. However, in the signal processing sense, their dynamic range is significantly lower.

By the way, there is also an equivalent to 24bit audio in photography: the 'raw' capture of most modern DSLR and other exchangeable-lens systems produces images with 12-16 bits. Not because those images are particularly 'nicer' to look at on their own, but mainly because they allow for significantly more headroom in terms of correcting/filtering the data.

Last edited by kalessin (25-06-2014 10:23)
Pianoteq 6 Standard (Steinway D&B, Grotrian, Petrof, Steingraeber, Bechstein, Blüthner, K2, YC5, U4, Kremsegg 1&2, Karsten, Electric, Hohner)

Reply by vjau

  • Avatar
  • vjau
  • Member
  • Offline
  • Registered: 08-11-2010
  • Posts: 46

Re: Bit depth, noise, dynamic range

In the way that I would use the terms, in terms of photography, a low resolution black and white digital image might have a range from white to black, but only have ten gray-scale steps between them. A higher resolution image might have the same range from white to black, but have thousands of gray-scale steps between them.

And how would you encode that ? Think of a numerical example and you will figure it out.

Reply by Jake Johnson

  • Avatar
  • Jake Johnson
  • Member
  • Offline
  • From: Atlanta
  • Registered: 04-05-2007
  • Posts: 4,501

Re: Bit depth, noise, dynamic range

kalessin wrote:
Jake Johnson wrote:
vjau wrote:

That's the same thing.
Number of possible amplitudes = dynamic range.
If you don't agree think photography.

No, no. Range indicates the lower and upper limits. Bit depth indicates the number of amplitude levels possible at a given instant. Very different things, aren't they?

EDIT: In the way that I would use the terms, in terms of photography, a low resolution black and white digital image might have a range from white to black, but only have ten gray-scale steps between them. A higher resolution image might have the same range from white to black, but have thousands of gray-scale steps between them.

vjau is actually right. The number of possible different values that can be present in a signal is exactly what dynamic range means. You are thinking of something different. I'll try to explain below.

Jake Johnson wrote:

Your original argument seemed to be that the main effect of a higher bit rate is an increase in dynamic range. I was saying that the more significant audible contribution was instead simply the number of amplitude levels simultaneously available.

I'll assume you mean bit depth and not rate. And yes, I said that: higher bit resolution means lower quantisation noise means more dynamic range, and I stand by that statement.

Yes, I meant bit depth, but you are still not understanding what I wrote. I never questioned anything you wrote about math or quantisation noise or dynamic range. I simply wrote that the ability to hear more amplitudes simultaneously is what made high bit rates desirable. I do not think that this was implied by noting that a higher bit rate creates a wider dynamic range.

Cheers. Let's move on to more important things. Like that Pianoteq thing.

Reply by custral

  • Avatar
  • custral
  • Member
  • Offline
  • Registered: 22-08-2010
  • Posts: 288

Re: Bit depth, noise, dynamic range

Say, Jake, don't think this an utter waste of time in a semantic sinkhole. For one thing, it's spaded up Monty (I donated BTW); and for another, this -

Jake Johnson wrote:

No, no. Range indicates the lower and upper limits. Bit depth indicates the number of amplitude levels possible at a given instant. Very different things, aren't they?

EDIT: In the way that I would use the terms, in terms of photography, a low resolution black and white digital image might have a range from white to black, but only have ten gray-scale steps between them. A higher resolution image might have the same range from white to black, but have thousands of gray-scale steps between them.

- quite accidentally comes close to a problem with manipulating photo histograms of image intensities. Trying to say how, long story involving populations of adjacent intensities, and it looks to offer a hint or so. Current manips raise the problems, and then say only So What about it. Better needed, and that must come about thru interpolating more (finer) grays into what must then be more pixels to house them, problem: where? Stark question, and the hint at offer is tag each of every histo-intensity's members with its (current-res) pixel address.

INSERT: Pixels within images each have the same x,y address for all of several bitplanes already, and this may be sufficient addressing, or bearing in mind the existing shortcomings, not - AI methods probably needed, in order not to buy more.

So thanks for the stark question. Good beginning, and from my perspective no waste of words in any way.

Last edited by custral (26-06-2014 08:14)

Reply by vjau

  • Avatar
  • vjau
  • Member
  • Offline
  • Registered: 08-11-2010
  • Posts: 46

Re: Bit depth, noise, dynamic range

Jake Johnson wrote:

Yes, I meant bit depth, but you are still not understanding what I wrote.

It's a shame you are taking it like that when you had the chance to learn something.

Reply by beakybird

  • Avatar
  • beakybird
  • Member
  • Offline
  • Registered: 29-03-2014
  • Posts: 209

Re: Bit depth, noise, dynamic range

vjau, you asked a rhetorical question: And how would you encode that ? Think of a numerical example and you will figure it out.

What is the answer?

Pianoteq 6 Std, Bluthner, Model B, Grotian, YC5, Hohner, Kremsegg #1, Electric Pianos. Roland FP-90, Windows 10 quad core, Xenyx Q802USB, Yamaha HS8 monitors, Audio Technica
ATH-M50x headphones.

Reply by vjau

  • Avatar
  • vjau
  • Member
  • Offline
  • Registered: 08-11-2010
  • Posts: 46

Re: Bit depth, noise, dynamic range

There is a confusion between the range of the display and the range of the data.
To display a pixel with the lowest value your screen is able to display and an other with the highest value he is able to display, you only need two values : 0 and 1. There are only two steps, hence it can be encoded with a one bit signal. The dynamic range is very narrow.
The fact that the screen is able to display a wide or a narrow gamut doesn't change that.

About the rhetorical example, if you have ten steps between pure black and pure white, pure black will be encoded as 0, and pure white as 11 (base 10). 11 decimal == 1011, you only need 4 bits, still low dynamic.

Now with "thousands" of steps, let's say 4096, you can encode your colors with 12 bits (2pow12=4096)

Reply by kalessin

  • Avatar
  • kalessin
  • Member
  • Offline
  • From: Germany
  • Registered: 06-12-2013
  • Posts: 306

Re: Bit depth, noise, dynamic range

vjau wrote:

About the rhetorical example, if you have ten steps between pure black and pure white, pure black will be encoded as 0, and pure white as 11 (base 10). 11 decimal == 1011, you only need 4 bits, still low dynamic.

Maybe this is a bit more clear: in that example, you will not see any details with a contrast of less than 1/11 of the gamut. In other words, you see almost no details, and that is what 'low dynamic range' means in this case. Or simplified further: when going from the complete gamut, how small a detail is still visible?

Last edited by kalessin (26-06-2014 23:33)
Pianoteq 6 Standard (Steinway D&B, Grotrian, Petrof, Steingraeber, Bechstein, Blüthner, K2, YC5, U4, Kremsegg 1&2, Karsten, Electric, Hohner)

Reply by Gilles

  • Avatar
  • Gilles
  • Member
  • Offline
  • From: Québec, QC
  • Registered: 18-07-2008
  • Posts: 3,344

Re: Bit depth, noise, dynamic range

To kalessin, and others interested, the statistical analysis for the 24 vs 16 bit audio test I wrote about is just out:

http://archimago.blogspot.ca/search?upd...results=26

...and you can rest assured, 16 bit is still statistically "sufficient", at least for this 140 person sample. Only 19 others gave the correct answer but one of the graphs shows that those giving 5 stars (high confidence) are split 50-50 in correctness !

I stand by my statement that I do hear a significant difference, but this may be due to more experience in analytical listening than average. For example I was a tiny bit less sure for the brass band test, of which I have no special accointance, than those for voice and piano.

Interesting read!

Reply by kalessin

  • Avatar
  • kalessin
  • Member
  • Offline
  • From: Germany
  • Registered: 06-12-2013
  • Posts: 306

Re: Bit depth, noise, dynamic range

Gilles,

I saw, and thanks. One question though: did you use a local randomiser like foobar2000's ABX plugin? The mind is a funny thing... I was almost certain I heard subtle differences, but then I used the ABX plugin: turned out I was not able to tell A from B at all once I did not know which file I was playing. The plugin gives you two choices (X, Y) and you have to guess which one corresponds to which input file. For each round it reshuffles and you select again. It would be very interesting if you really can reliably pick out the 24bit samples round after round even with that addin.

Last edited by kalessin (27-06-2014 22:00)
Pianoteq 6 Standard (Steinway D&B, Grotrian, Petrof, Steingraeber, Bechstein, Blüthner, K2, YC5, U4, Kremsegg 1&2, Karsten, Electric, Hohner)

Reply by Gilles

  • Avatar
  • Gilles
  • Member
  • Offline
  • From: Québec, QC
  • Registered: 18-07-2008
  • Posts: 3,344

Re: Bit depth, noise, dynamic range

kalessin wrote:

Gilles,

I saw, and thanks. One question though: did you use a local randomiser like foobar2000's ABX plugin? The mind is a funny thing... I was almost certain I heard subtle differences, but then I used the ABX plugin: turned out I was not able to tell A from B at all once I did not know which file I was playing. The plugin gives you two choices (X, Y) and you have to guess which one corresponds to which input file. For each round it reshuffles and you select again. It would be very interesting if you really can reliably pick out the 24bit samples round after round even with that addin.

No, I did my own randomizing, meaning I didn't look at the filename for a number of compared hearings, and kept looking for that special difference that I isolated, until I was sure which was which. When I heard for instance that the singer stood out from the accompaniment, I could reliably know which was which. I looked for things like that in the piano also, moving notes on the soundboard as I said, and also, curiously, the pedal sound was more present in the 24 bit version. The brass example is mostly in details of the die-out sound, but I was not that sure for this one (4 stars instead of 5)

The blogger insisted at first that maybe the longer dying-out of sound would give a hint, having a theoretically lower noise floor with 24 bit, but I didn't hear much difference there. Like I said, what I hear is mostly details that to me are a bit masked in the 16 bit version.