Jitter Correlation to Audibility

[stv014]

  I inverted A and compared A(6ps)  vs B  . Most of the time the two files are very close but at higher frequencies they diverge drastically as below

 
I only found a small difference: mostly under -90 dBFS, except for some peaks, but it is again much higher than it should be for 6 ps jitter.

 

[UltMusicSnob]

Okay, I'm going to just leave Crane Song alone for now, until I understand what's up with their files.
 
In the meantime, path30n versus the latest, path30jr:
 
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/26 20:14:38
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\path30n.flac
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\path30jr.flac
20:14:38 : Test started.
20:15:05 : 01/01  50.0%
20:15:10 : 02/02  25.0%
20:15:15 : 02/03  50.0%
20:15:25 : 03/04  31.3%
20:15:32 : 04/05  18.8%
20:15:39 : 05/06  10.9%
20:15:49 : 05/07  22.7%
20:16:00 : 06/08  14.5%
20:16:14 : 07/09  9.0%
20:16:30 : 08/10  5.5%
20:16:40 : 09/11  3.3%
20:17:23 : 10/12  1.9%
20:17:26 : Test finished.
 ----------
Total: 10/12 (1.9%)
 
Like all the tests, this ABX requires replication, including multiple testers, before there's any point drawing conclusions about it. Getting into a pretty subtle area here. The earlier files seemed to be close together in terms of the audible effect of jitter, but path30jr is more challenging. Still using the same cue, quality of the piano attack in soft passages, but there's less of it here.

 

[Don Hills]

   
How is that relevant to my suggestion to UltMusicSnob to fix the inverted file before an ABX test ?

 
Only relevant to the fact that it needs to be done. I was referring to the total of the comments, I shouldn't have quoted your post specifically. Sorry.

 

[stv014]

Originally Posted by UltMusicSnob /img/forum/go_quote.gif
 
In the meantime, path30n versus the latest, path30jr:

 
That is a really excellent result, the difference between the files is always less than -90.3 dBFS (= 1 LSB in 16-bit resolution), except for a low level "click" at the end of the file, and only rarely has peaks above -96 dBFS.

 
With path30j.flac, you should now be able to get a 20/20.

 

[nick_charles]

   
That is a really excellent result, the difference between the files is always less than -90.3 dBFS (= 1 LSB in 16-bit resolution), except for a low level "click" at the end of the file, and only rarely has peaks above -96 dBFS.
 
With path30j.flac, you should now be able to get a 20/20.

 
Astonishing would be a better word, unless UMS is cranking the volume to insane levels it is getting close to the edge of redefining (at least for UMS)  audibility thresholds. Can you graph a short segment with and without the reduced jitter, it would be interesting to see if there are any added sidebands that may be being detected, if it is just noise that UMS is picking up then it is remarkable.

 

[nick_charles]

  Okay, I'm going to just leave Crane Song alone for now, until I understand what's up with their files.
 
In the meantime, path30n versus the latest, path30jr:
 
 
20:17:23 : 10/12  1.9%
 

 
 
Is this at normal listening levels ?

 

[mikeaj]

  Astonishing would be a better word, unless UMS is cranking the volume to insane levels it is getting close to the edge of redefining (at least for UMS)  audibility thresholds. Can you graph a short segment with and without the reduced jitter, it would be interesting to see if there are any added sidebands that may be being detected, if it is just noise that UMS is picking up then it is remarkable.

 
I forget (and haven't been following much)... the difference here isn't uncorrelated noise, is it? I mean, depending on what kind of "difference" signal we're looking at, it may or may not be easy to distinguish from background noise even if at a lower level than the noise. The bigger deal is the level relative to the signal power.
 
Maybe one thing would be to look at is some kind of noise at a similar power level but uncorrelated. That would help distinguish between the level of the disturbance and its nature (which is causing the audibility).
 
Anyway, impressive for sure. And it's very likely that I am misunderstanding something because of lack of knowledge and/or not really following developments.

 

[bigshot]

There is one control that hasn't been applied here.

 

[nick_charles]

   
I forget (and haven't been following much)... the difference here isn't uncorrelated noise, is it? I mean, depending on what kind of "difference" signal we're looking at, it may or may not be easy to distinguish from background noise even if at a lower level than the noise. The bigger deal is the level relative to the signal power.
 
Maybe one thing would be to look at is some kind of noise at a similar power level but uncorrelated. That would help distinguish between the level of the disturbance and its nature (which is causing the audibility).
 
Anyway, impressive for sure. And it's very likely that I am misunderstanding something because of lack of knowledge and/or not really following developments.

 
There are two components, there is uncorrelated noise and there are signal-correlated distortion sidebands, STV014 I think can put either or both in as required, the current sample has both types. The current samples do not have random noise. The distortion sidebands are potentially more audible but that depends on the level they are at, which is low. For reference a good (but not outrageous) DAC will have jitter sidebands no higher than -110db and a very good DAC but still affordable will have sidebands below -125db, the $249 Dragonfly manages this as does the $379 Musical Fidelity V-DAC. 

 

[nick_charles]

  There is one control that hasn't been applied here.

 
? Do tell !

 

[stv014]

  There are two components, there is uncorrelated noise and there are signal-correlated distortion sidebands, STV014 I think can put either or both in as required, the current sample has both types. The random noise effectively lowers the bit-depth from 16 down to about 15 bits, this is hard to detect at normal listening levels.

 
Uncorrelated noise is not included in the test samples, that is, a silent input would produce a silent output (other than quantization noise). It is frequency modulation by a mix of lowpass filtered noise and sine waves. So, the amount of noise/distortion in the output file is proportional to both the amplitude and frequency (or, in other words, the slew rate) of the input signal.

 

[stv014]

Can you graph a short segment with and without the reduced jitter, it would be interesting to see if there are any added sidebands that may be being detected, if it is just noise that UMS is picking up then it is remarkable.

 
I will try finding a segment where it is visible, but I am not sure if it will be easy. Most of the time the sidebands will probably be masked by "noise" in the input signal (music often does not have a very high dynamic range within an FFT frame).

 

[nick_charles]

I took the opening chord and analyzed that (about 3.4 seconds)  0.773 to 4.222 seconds
 

Using a 16K FFT in Audacity I plotted the differences as above. The biggest negative differences were as follows
 

Frequency (Hz)	No Jitter	reduced jitter	diff
4359.4	-98.243393	-98.141502	-0.10189
6843.8	-113.311104	-113.223793	-0.08731
375.0	-64.395424	-64.319214	-0.07621
5906.3	-109.054237	-108.983971	-0.07027
2929.7	-95.858932	-95.790863	-0.06807
4986.3	-112.535103	-112.470146	-0.06496
4388.7	-96.874069	-96.809158	-0.06491
5285.2	-107.907715	-107.843712	-0.064
4037.1	-110.456573	-110.393379	-0.06319
638.7	-72.387726	-72.324638	-0.06309

at the other end the biggest positive differences

frequency (Hz)

No Jitter

reduced jitter

diff

2595.7	-88.965973	-89.008209	0.042236
3246.1	-81.961166	-82.003761	0.042595
3152.3	-103.288048	-103.331123	0.043075
1546.9	-84.732315	-84.777336	0.045021
1031.3	-82.501305	-82.547699	0.046394
4951.2	-108.961998	-109.010223	0.048225
4031.3	-105.531876	-105.583542	0.051666
3755.9	-100.858963	-100.911011	0.052048
3527.3	-93.297768	-93.351547	0.053779
3357.4	-105.328659	-105.384407	0.055748
4763.7	-112.782753	-112.839233	0.05648
4394.5	-99.473022	-99.538902	0.06588
4002.0	-109.411629	-109.477615	0.065986
1605.5	-84.671112	-84.738846	0.067734
3960.9	-108.063446	-108.138374	0.074928

 
 
Average difference = -0.0021 db
 
These differences really should not be audible at least not for this segment.
 
here is my spreadsheet for those interested http://www.divshare.com/download/24569132-4d7

 

[bigshot]

Not even in the same universe as audible.

 

[UltMusicSnob]

   
 
Is this at normal listening levels ?

Yes, these are normal listening levels, it interferes with identification if I crank levels. The cue is the moment of attack, as the hammers strike the strings on quiet chord. The loud chord at the beginning is too complex, at least with what I'm able to do so far. I'm still cuing on "focus", which is a poorly defined descriptor. The jittered chord also sounds different in a way I might call frequency balance--it becomes *slightly* less realistic than the 'n' version, out of balance in its frequency spectra as I'm listening for the full round piano tone.