Seeing how you go to shows and exhibit all the symptoms of subjectivism with modded players, ribbon speaker cables, etc, I say your are in denial and hence the reason I asked about other symptoms.

And here is the reference in Dr. Toole's Book

I don't know Ammar.  Maybe I need the snake oil flat ribbon speaker cable you use with your speakers to hear them.

I provided my article to you on this topic explaining the same: http://www.madronadigital.com/Library/RoomDynamicRange.html.  And noted the reference at the end:

“Noise: Methods for Estimating Detectability and Threshold,” Stuart, J. Robert, JAES Volume 42 Issue 3 pp. 124-140; March 1994

I think there's some confusion because the UEN (red curve in the graph quoted many times in this thread) is quoted for 1Hz wide bands, and the noise floor in the same graph is quote for 1Hz wide bands. That's fine. The noise will cross the UEN at the same real-world signal level that a one-third octave measure of the same noise would cross the regular (tone referenced) threshold of hearing curve. The 1Hz wide measures are obviously very different from the one-third octave measures, but as long as you compare like with like, you get the same threshold for noise in the range that one-third of an octave is a regional approximation to the width of a critical band (i.e. the filter in the ear). Where it's not, it doesn't work, and an accurate UEN is a better way of doing it than one-third octave measures.

If you don't know what you're doing, it goes "wrong" for tones, or very shaped noise. The fact the curve for tones ends up being 20dB+ higher than the curve for noise shows part of the problem. The fact that a 1Hz bin implies an analysis time of one second is the other problem: any music which changes spectrum within a second isn't being analysed properly.

I'm not suggesting any mistake is being made in the paper (though as ever, I'd like more details). I do think people are getting a bit carried away in their analysis, claiming things as fact that they possibly don't fully understand.

Page 10: This could be an effect of quantization distortion; it is interesting that this was audible even in a 24-bit system, and is consistent with the hypotheses of Stuart [29] that 16 bits are not sufficient for inaudible quantization.

It is possible that that these two qualitative effects on sound quality cancelled out each other for the conditions where 16-bit quantization plus RPDF dither was applied, which could explain why these signals were harder to discriminate from their unfiltered counterparts.

I'm still wondering if the paper doesn't contradict itself by predicting in Fig. 3 that the dither should be inaudible and at the same time finding in the tests that dither made a significant audible difference.

Another thing that bothers me is that the paper uses the term "quantization" for what seems to be truncation (undithered wordlength reduction). Please correct me if I'm wrong.

Are they masking truncation artifacts with dither noise instead of properly (RPDF is better than nothing) dithering before truncation ?

I'd like to see 24-bit quantization noise plotted in Fig.3. It should be far below the thresholds of human hearing, so why would this be audible, even in a 24-bit system ?

Quote from: Kees de Visser on 2014-11-29 15:39:39

I'm still wondering if the paper doesn't contradict itself by predicting in Fig. 3 that the dither should be inaudible and at the same time finding in the tests that dither made a significant audible difference.

Yeah, assuming this is not a result of chance or bias in their methodology or something like that, the results show that for 22.05 kHz cutoff condition 1 (24 bit TPDF) was easier to distinguish from the original than both condition 2 (16-bit "quantization") and condition 3 (16-bit "quantization and RPDF").
So their claim that 16-bit made a significant audible difference seems to be refuted by their own results.

They really should have called it "16-bit truncation" instead of just "16-bit quantization".
And their phrasing "16-bit quantization and rectangular dither" really does sound like adding dither after truncation, but I doubt that they would make such mistakes. At least I hope it's just bad phrasing.

Why not just use TPDF? Why the weird phrasing? Why do they mix up quantization distortion and dither noise, and audibility of these things?


edit: Actually, the anecdotal subjective impressions would make more sense if they screwed up quantization.

The results they achieved were IME close enough to random guessing to not totally eliminate the possibility that they were  in fact random guessing.  56% right after working as hard as they did is really pretty disappointing.

Another point is that best practices says that the dither should be TPDF and have perceptually shaped PSD.  They had two opportunities to use best practices and blew them both off. Pretty ironic given that Meridian used to promote the benefits of equipment they sold that did both things right.

TPDF dither was used

You are wrong . In the context of peers reading such a paper, saying that going from 24 bits to 16 bit is quantization is correct terminology.  A more precise one would be re-quantization but that "re" is understood.

As it happens I had another AES paper open which directly addresses this:
Subtractive Dither for Internet Audio
BEN DENCKLA

Dithered quantization can be useful in preparing audio files for distribution on the Internet, since this often involves quantizing audio down from its original word length (typically 16 bits) to 8 bit in order to achieve a form of lossy compression, for example, offering 2:1 compression of a 16-bit source.  This is sometimes referred to as “requantization” to distinguish it from the quantization of a truly continuous source, such as an analog electronic waveform, but we will not bother with that distinction here.

So the terminology is fine.

No, the context in that quote is the combination of filtering artifacts and requantization combining in some way to generate that listening test result.

I'd like to see 24-bit quantization noise plotted in Fig.3. It should be far below the thresholds of human hearing, so why would this be audible, even in a 24-bit system ?

The results they achieved were IME close enough to random guessing to not totally eliminate the possibility that they were  in fact random guessing.  56% right after working as hard as they did is really pretty disappointing.

Another point is that best practices says that the dither should be TPDF and have perceptually shaped PSD.  They had two opportunities to use best practices and blew them both off. Pretty ironic given that Meridian used to promote the benefits of equipment they sold that did both things right.

Or are you seriously suggesting that 16-bit RPDF cancels out the alleged effects of the lowpass filters?

Quote from: amirm on 2014-11-29 17:46:08

TPDF dither was used

Yep, by the BS crew...and here is the result:

Quote from: ajinfla on 2014-11-29 18:17:59

Quote from: amirm on 2014-11-29 17:46:08

TPDF dither was used

Yep, by the BS crew...and here is the result:

This gets so tedius!

From:
The audibility of typical digital audio filters in a high-delity playback system  AES paper 9174

Page 5

"
After filltering with either FIR filter, the signals
were either unchanged or were quantized to 16-
bit. The quantization either included RPDF (rect-
angular probability density function) dither or did
not.

Page 8

"A within-subjects ANOVA was performed on the
data for all eight listeners with two factors: filter
cutoff frequency (low or high) and quantization
type (none, 16-bit, or 16-bit plus RPDF dither).

Page 9:

"
Post-hoc Fisher tests based on a least-signicant dif-
ference of means at a 5% level of 0.1394 for quanti-
zation showed that performance was worse for 16-bit
quantization plus RPDF dither (mean=1.773) than
for no quantisation (mean=1.902) or 16-bit quanti-
sation alone (mean=1.909).

Page 9-10

"Every condition where 16-bit quantization was ap-
plied gave performance that was signicantly better
than chance. Performance was signicantly worse
for 16-bit quantization plus RDPF dither than for
no quantisation or 16-bit quantization alone. This
suggests that the eect of adding the RPDF dither
on top of the 16-bit quantization and FIR ltering
was to make it more dicult to identify that process-
ing had been applied to the signal, which is perhaps
counterintuitive. To try to explain this nding, we
turn to subjective descriptions of what listeners de-
scribed hearing in these tests.
"

Page 10

"It is possible that that these two qualitative ef-
fects on sound quality cancelled out each other for
the conditions where 16-bit quantization plus RPDF
dither was applied, which could explain why these
signals were harder to discriminate from their unl-
tered counterparts."

"All forms of processing tested here were audi-
ble, except for one condition where performance was
signicantly dierent from chance at the 6.7% level,
including emulated downsampling lters at standard
sample rates and 16-bit quantization with or without
RPDF dither. Dierences were demonstrated here
in a double-blind test using non-expert listeners who
received minimal training.

"16-bit quantization with and without RPDF
dither can have a deleterious eect on the listen-
ing experience in a wideband playback system.

That is what they are hypothesizing to explain the outcome in the article.  That is what the topic was: what the text meant.  And I explained it.  So no, it is not my suggestion.

Once again, the filter test used the above TPDF.  Per that same quote, it rules out audibility of quantization distortion in the interim math used for the filter.

The interesting questions are:
What is the crossover frequency for the tweeter?
How much power does it receive during loud passages?
Where are the measurements for this tweeter?

Once again, the filter test used the above TPDF

This gets so tedius!

You telling me?  Here is the relevant part:

[...]

That processing, filtering down to half bandwidth of 44 Khz, while maintaining 24 bit sample resolution was done with TPDF.  The results were positive. 

Since any down conversion first requires a filter, then the above demonstrates that the process is not transparent according to this test data.

Quote from: amirm on 2014-11-29 17:24:31

As it happens I had another AES paper open which directly addresses this:
Subtractive Dither for Internet Audio
BEN DENCKLA

Dithered quantization can be useful in preparing audio files for distribution on the Internet, since this often involves quantizing audio down from its original word length (typically 16 bits) to 8 bit in order to achieve a form of lossy compression, for example, offering 2:1 compression of a 16-bit source.  This is sometimes referred to as “requantization” to distinguish it from the quantization of a truly continuous source, such as an analog electronic waveform, but we will not bother with that distinction here.

So the terminology is fine.

No, you are stepping in exactly the trap that Kees de Visser pointed out, confusing truncation with dithered quantization.

I'm still wondering if the paper doesn't contradict itself by predicting in Fig. 3 that the dither should be inaudible and at the same time finding in the tests that dither made a significant audible difference.

Dither is an optional processing step in quantization.

Quote

I'd like to see 24-bit quantization noise plotted in Fig.3. It should be far below the thresholds of human hearing, so why would this be audible, even in a 24-bit system ?

The quantization in 24 bit domain used TPDF post output of the filter. You are asking for level of that?

Listeners described that quantization gave a “roughness” or “edginess” to the tone of the instruments, and that quantization had a significant impact on decay, particularly after homophonic chords, where “decay was sustained louder for longer and then died suddenly”. This could be an effect of quantization distortion; it is interesting that this was audible even in a 24-bit system, and is consistent with the hypotheses of Stuart [29] that 16 bits are not sufficient for inaudible quantization.