Distortion Surface: old measurement, new approach

The red dot depicts where a particular "scientist" inappropriately tests for SINAD to judge amplifiers.
The black square encompasses the typical operating area where many headphones produce between 75 and 90 dB SPL.





To be very clear:
2Vrms = 8.23 dBu
The particular THD=N (SINAD) test referenced is a single operating point measurement:
2Vrms @ 1 KHz

 

I'd up my patreon donation for that.

 

So If I read these plots correctly, the DSHA-3F has more relative distortion in the sub 100Hz region than the iFi Diablo for headphone SPL between 75 and 90db? And, the overall distortion of the DSHA-3F would be "better" at higher dBu levels?

If I understand these plots, It is curious that the AmirNAD does suggest in this particular scenario that the Diablo would be amp with less distortion. What would be a real world scenario for this plot? Would this suggest that a transducer under these conditions could have relatively higher distortion in the mid and lower bass with the DSHA-3F?

This is fascinating stuff. Thanks @atomicbob

 

The plots were created with default automatic settings which spreads the colors over the range of the data rather than a fixed level. I'm working through the program to see if I can fix the range and colors such that the plots would be consistent across devices measured. Currently each device has its own color map which makes comparisons tricky as mentioned by @Wilewarer .

This visualization is by no means mature yet and requires a lot of manual effort. And keep in mind that so far I have been showing only THD+N contours which are still oversimplifications. When I can automate the process better it will be far more informative to look at D2, D3 and 4+HD+N contours.

 

Hey guy's, just joined to say a big thank you to AtomicBob & all the other great contributions to this thread
I do ( VERY begrudgingly) have a look at a few measurements on ASR & although I know they are limited & rushed (a reflection on the guy doing the measurements sadly), it's good to see some objective stuff.
What really pis**s me off is the ridiculous refusal to understand the measurements should explain what we hear & not suggest what we should hear.
All the silly 'if you don't agree' shite, 'it's obviously your preference' is really tedious.
Thank you all for trying to advance our actual understanding of what our perception actually looks like

 

I wholeheartedly agree that the Pink Panther is an idol for the Temple of SINAD zealots. I didn't mean to infer otherwise. Nor do I think I could hear distortion at -55db at 100Hz.

I'm trying to wrap my brain around the plots. Thanks for clarifying.

 

Would be interesting to test a SE tube amp, I suspect it would be the the exact opposite of the IFI, with distortion increasing with volume rather than decreasing. I guess if it was transformer based you'd also see distortion increases in the bass regions.

Would also be interesting to see how the various harmonic distortions are represented across the volume curve, my simulating of a lot of tube circuits seems to imply that while distortion increases with volume, the 2nd Harmonic becomes more dominant as volume increases.

 

Ok, I think I have a reasonable template for future contour visualizations. Here are the same two contour maps seen many times above with color map fixed scales such that the two may be directly compared:





This is still a very manual process. As I automate both the dScope sweep measurement data capture and hopefully can template the visualization such that I need only import new data sets from various components, more components will have such distortion visualizations presented. But for now I must return to finishing iFi iDSD Diablo technical measurement series and post both it and KTE Spring 3 data. Very behind.

Choice of iFi Diablo and DSHA-3F for visualization experimentation occurred because they have both been on the lab bench for some time.

 

Here is a quick example of what you could do with gnuplot:



The above image was created out of the "DSHA-3F THD+N vs Amplitude" sweep files. "Bin" is not necessary, so I removed that and the comment column. Each file in your set represents an isoline to the program, so I line-separated each set in the input (see "sample.dat"). Left-to-right, you have Frequency (which was not in your files, it was in the name of the file), Level, and Distortion.

The contour plot implements a few of suggestions made in thread: (1) how to fix the scale for any input (here shown, -100 to -40 for distortion, with -50 not-so-bad-red), and (2) the axes as people like to see them. I even left space to the right for your avatar

The gnuplot commands are as follows:

Code:

set size square # The output image will be square
set title 'DSHA-3F THD+N vs Amplitude'
set pm3d map
set pm3d interpolate 0,0 # Smooth!
set logscale x 10 # Sets x axis to be log10
set xrange [50:10000] # Frequency min/max
set yrange [-30:10] # Level min/max
set cbrange [-100:-40] # Distortion min/max
set palette maxcolors 128 # Colors to be used for contour
set palette rgb 33,13,10 # Palette - blue to red
set xlabel 'Frequency (Hz)'
set ylabel 'Level (dBu)'
set cblabel 'Distortion (dB)'
splot 'sample.dat' # File (see format if you want to automate)

The input file:

Code:

50    -30    -48.51479815
50    -28    -47.74858161
50    -26    -47.38059715
50    -24    -47.47525402
50    -22    -48.04836509
50    -20    -49.03736438
50    -18    -49.99591043
50    -16    -51.23152167
50    -14    -52.60510213
50    -12    -54.0152624
50    -10    -55.40206164
50    -8    -56.65650518
50    -6    -57.87432279
50    -4    -59.24633334
50    -2    -57.18654718
50    0    -56.75745403
50    2    -56.48788358
50    4    -55.72566967
50    6    -55.43529034
50    8    -54.65228057
50    10    -53.92999866
 
100    -30    -59.89135551
100    -28    -59.86837688
100    -26    -57.52696806
100    -24    -56.22405152
100    -22    -55.35803236
100    -20    -54.96147586
100    -18    -55.04115706
100    -16    -55.68650288
100    -14    -56.51233777
100    -12    -57.48859502
100    -10    -58.62781461
100    -8    -59.84965092
100    -6    -61.1913031
100    -4    -62.45204197
100    -2    -62.76842082
100    0    -62.66055872
100    2    -61.71304081
100    4    -60.98641783
100    6    -60.29687346
100    8    -59.29387378
100    10    -57.95342954
 
500    -30    -60.33623605
500    -28    -62.18568117
500    -26    -64.28832067
500    -24    -66.03018054
500    -22    -68.08766237
500    -20    -70.23228748
500    -18    -72.29692042
500    -16    -74.11572427
500    -14    -74.92185923
500    -12    -74.28957691
500    -10    -73.07816127
500    -8    -72.30038032
500    -6    -71.85548543
500    -4    -72.2509223
500    -2    -72.78784767
500    0    -73.68648935
500    2    -74.66281453
500    4    -75.67402534
500    6    -76.75559422
500    8    -77.61575289
500    10    -78.23267602
 
1000    -30    -60.34501535
1000    -28    -62.46392228
1000    -26    -64.15739909
1000    -24    -66.21813687
1000    -22    -68.1544676
1000    -20    -70.34821644
1000    -18    -72.21987075
1000    -16    -74.27940442
1000    -14    -75.70231126
1000    -12    -78.05092004
1000    -10    -80.11954597
1000    -8    -81.28357202
1000    -6    -81.50255882
1000    -4    -80.51319829
1000    -2    -79.76956866
1000    0    -79.33508855
1000    2    -79.58771539
1000    4    -80.16713476
1000    6    -81.05159297
1000    8    -81.56148062
1000    10    -81.82427992
 
5000    -30    -60.15360923
5000    -28    -62.40522227
5000    -26    -63.93379879
5000    -24    -65.91176666
5000    -22    -68.24585106
5000    -20    -69.93239776
5000    -18    -72.35688333
5000    -16    -74.49378262
5000    -14    -75.92106051
5000    -12    -78.6050384
5000    -10    -79.99937262
5000    -8    -81.99965654
5000    -6    -84.10073114
5000    -4    -85.94936673
5000    -2    -87.85180537
5000    0    -89.93172999
5000    2    -90.73948094
5000    4    -93.4042793
5000    6    -95.66991653
5000    8    -96.92917294
5000    10    -97.78605221

Edit: fixed data source snafu (Thanks, @Biodegraded!)

 

Looks nice, but looking at Bob's graph and the version of his data I have, I think your 500 Hz column is off - it should be in the -60 t0 -78 dB range, not down below -87. Edit: fix noted

I must learn GNUPlot. I'm guessing it'd be easier with a GUI, e.g. JAGPI.

 

Nice catch! I copied the wrong file it seems. Will fix and edit post. Tnx!

Edit: Now fixed code, data, and image.

 

You know how we have compensation curves for raw FR? I wonder if we could do something similar with distortion.

Maybe not so much with these surface plots, but with stuff like how 2nd order generally sounds benign while 3rd order is harsher, etc. I suppose it would be more obvious on the individual harmonics charts, and maybe consolidated into a THD/whatever plot maybe that would alter data into new trends or we'd see certain areas pop out more. Of course, people would then fight over the compensation curves, but hey it's a nugget of thought.

 

It is very hard to reconstruct a signal after unknown noise has been applied. There are some promising AI-driven recent developments on this topic:

https://www.researchgate.net/public...n_Effects_in_Music_Using_Deep_Neural_Networks

 

Maybe try Matplotlib instead?

Matlab has nice color maps (jet or parula are both easy to read).

 

Oh man, I haven’t had a Matlab license since… hmm.. Macs had PowerPC chips…

but yeah, choices abound.

Regardless of plotting software and cost, I think agreeing on what we look at for comparisons (fixed axis, etc), what we call “bad”, etc is going to be very important moving forward. We can automate anything.

The other thing that would be useful is to set up a GitHub or something for the community to share datasets, collection parameters and methodologies, and plotting scripts. Reproducibility, after all, is a desirable property in science. If there is interest (especially from data generators) I can help.

 

@atomicbob awesome as always

Anyone here have a SAS JMP license?

 

I had a couple of play hours this afternoon/evening and I set myself to reboot my very rusty knowledge of Python's matplotlib. More work to do (I didn't bother to place legends yet), but I'm now in control of what needs to be done to create comparable visualizations. Here's what it looks like, I can tweak as needed.

 

There's the BBC weighted harmonic distortion as described here:
https://superbestaudiofriends.org/i...y-51-technical-measurements.8387/#post-274738

 

I assume its simply a difference in graphing programs, but I'm surprised that @atomicbob 's graph appears to show different results than the others when all three are using the same data.

Eg.,
Atomicbob's has a well defined low distortion (~-100db) area that appears centered near -5dbu and 12000Hz 3000Hz... that area appears to be missing on the other two plots. Are the other two programs smoothing the plot too much? or is the software used in the first plot too coarse?

------edit------
because graphs confuse me