Succesful ABX of 24/96 vs. 16/44.1 Options

[googlebot]

The maximum sample rate of my digital output is 96 kHz at 24 bit. To eliminate as many variables as possible I locked the output to that format and pre-converted the test files accordingly using the highest quality options available. As source material I have used the 24bit/192kHz track from this thread.

The high resolution contestant (24/96) was produced using the following settings:

CODE

sox Sample_192000Hz_24bit.wav 24-96000.wav rate -v 96000 dither -s

That is the setting recommended by Sox' developers for SRC with large target rates. Basically linear phase filters and sloped TPDF dither.

For the Redbook contestant, the 24bit/192khz track was first down-sampled to 16bit/441000Hz:

CODE

sox Sample_192000Hz_24bit.wav -b 16 16-44100.wav rate -v 44100 dither -s

And then up-sampled to 24bit/96kHz, again at the highest quality settings:

CODE

sox 16-44100.wav -b 24 16-44100-24-96000.wav rate -v 96000 dither -s

My hearing tops at about 17 kHz. Still I can hear pretty clear difference over a pair of Elac FS 607 X-Jet speakers, which are rated for 28-50000Hz (IEC 268-5):

CODE

foo_abx 1.3.4 report
foobar2000 v1.0.3
2010/06/07 18:25:51

File A: C:\Dokumente und Einstellungen\Christian\Desktop\16-441to24-96.wav
File B: C:\Dokumente und Einstellungen\Christian\Desktop\24-96.wav

18:25:51 : Test started.
18:26:44 : 01/01  50.0%
18:27:09 : 02/02  25.0%
18:27:26 : 03/03  12.5%
18:27:44 : 04/04  6.3%
18:28:00 : 05/05  3.1%
18:28:21 : 06/06  1.6%
18:28:38 : 07/07  0.8%
18:28:52 : 08/08  0.4%
18:29:06 : 09/09  0.2%
18:29:17 : 10/10  0.1%
18:29:33 : 11/11  0.0%
18:29:46 : 12/12  0.0%
18:30:03 : 13/13  0.0%
18:30:26 : 14/14  0.0%
18:30:42 : 15/15  0.0%
18:31:04 : 16/16  0.0%
18:31:08 : Test finished.

----------
Total: 16/16 (0.0%)

I do not think, that I actually hear the presence or absence of stationary HF content. But what I am hearing is louder and somewhat smeared sounding transients in the Redbook version. That is exactly the type of pass-band artifact to be expected from low-pass filtering, but until lately I hadn't considered that to be audible, when done right. And since that step is necessary it can also not be avoided for Redbook delivery.

I do not hear a difference over my regular Canton speakers, which top at about 20kHz! Neither over my Grado headphones. Since both act as a mechanical low pass, I think they just cause the same artifacts in the pass band as the digital low-passing.

Until today I have mostly laughed at high resolution apologists. But that ABX result somewhat changed my perspective. Maybe those, who say that you do need excellent speakers to hear a difference are actually right? On the other hand this is just one very artificial and synthetic sample. The results can very likely not be extrapolated to real world recordings. I don't know. What do you all think?

The speakers are here for testing right now, but I think I'm going to buy them.

PS: I have uploaded the samples.

This post has been edited by googlebot: Jun 7 2010, 18:36

[greynol]

Have you tried other resampling algorithms?

After looking at the difference between the two files in time and frequency domain, I'm not terribly surprised you can tell the difference.
EDIT: Scrap that. Frequency vs. amplitude as well as a spectral plots shows that the differences are pushed well into the are you claim not to be able to hear.

This post has been edited by greynol: Jun 7 2010, 20:31

[googlebot]

I really can't explain it, yet. Maybe something's wrong with the speakers or amplifier. There should be several HA regulars, who work in the field of science and have access to high quality (maybe also high bandwidth) listening setups. Maybe those can give the sample a try.

This post has been edited by googlebot: Jun 7 2010, 19:29

[greynol]

I don't have a good set of headphones or a quiet enough environment to think I'd even stand a chance.

I can say that I have listened to them and would certainly not dismiss the sample as being too synthetic or artificial.

[IgorC]

CODE

foo_abx 1.3.4 report
foobar2000 v1.0.3
2010/06/07 16:14:52

File A: C:\Documents and Settings\Administrador\Mis documentos\Downloads\24_96.flac
File B: C:\Documents and Settings\Administrador\Mis documentos\Downloads\16_441to24_96.flac

16:14:52 : Test started.
16:17:04 : 01/01  50.0%
16:17:31 : 02/02  25.0%
16:18:20 : 03/03  12.5%
16:19:20 : 04/04  6.3%
16:20:51 : 05/05  3.1%
16:20:52 : Test finished.

----------
Total: 5/5 (3.1%)

Headphones Sennheiser HD 447. Soundcard Audigy SE 24/96.
foobar -> output -> 24 bits

[Cavaille]

I´d love to do it but right now I´m suffering from the effects of a bad cold I had last week, my hearing is very impaired - so I can´t decide between a violin and an organ

This topic is very fascinating. Thank you for the time you spent with it. Have you tried several other resamplers also? They may have differing low-pass-filtering algorithms. I know that Sox has a very good reputation but there are others equally high rated (SSRC for example). Or maybe worse resamplers? Like Sound Forges? Or even iZotope RX Advanced where you can modify the slope of the low-pass-filtering? This test seems to be easily repeatable so the results could be very interesting.

[udauda]

16_44to24_96.wav has an unusual HF noise component(peak level is around -77dBFS) in the 16~21kHz range. (possibly due the effect of dithering?) Well whether the noise is audible or not, it seems there's already someone who successfully ABXed the original samples.

This post has been edited by udauda: Jun 8 2010, 02:23

[googlebot]

16_44to24_96.wav has an unusual HF noise component(peak level is around -77dBFS) in the 16~21kHz range. (possibly due the effect of dithering?)

That is indeed noise shaped dither. Sox's -s switch defaults to Shibata for Redbook, which quite aggressively mirrors the human ATH. Perceptually that should be the best choice. The higher resolution only gets sloped TPDF, because Shibata isn't available for those rates. I could have gone all-TPDF, but it was my intention to give the 16 bit sample the best conversion possible.

Well whether the noise is audible or not, it seems there's already someone who successfully ABXed the original samples.

Strictly speaking, there are no "original samples". There is just one original 192kHz sample and there are several ways to produce a Redbook version from that. There are also several ways to mess the latter up and that's why I have documented each step in detail.

This post has been edited by googlebot: Jun 8 2010, 04:01

[greynol]

it seems there's already someone who successfully ABXed the original samples.

That was a second pair of samples IgorC ABXed. He's been busy (or has had some spare time?).

[JSW]

Well, I think it would be a good idea to repeat the test with a flat-dithered 16/44.1 version of the file, or another test to separate the possible effects of static HF noise from the noise-shaping from the effect of the low-pass filter. I think that flat TPDF is the best possible choice for conversion to 16/44.1.

[Eudaimonius]

Could this be related to what's described in this quote, which I thought was just marketing blurb?

Another reason why ribbon tweeters are so supremely accurate is their ability to reproduce frequencies far above our normal hearing range. Even though we can’t hear these “ultrasonics,” the lower harmonics they generate do affect the way we experience the stereo image. For instance, a 30 kHz signal is too high for us to hear, but we can hear its first sub-harmonic (15 kHz), second sub-harmonic (7.5 kHz), and third (3.75 kHz), etc. All these sub-harmonics combine with the primary sound to create stunning realism.

[Arnold B. Kruege...]

googlebot:

I don't see anything obvioiusly wrong with the samples.

How sure are you that the listening environment that produces positive results does not have enough IM to cause problems?

You might want to exclude nonlinear distortion in the playback system as a potential source of problems with the test by doing some listening tests with test files composed of sine waves 4 KHz apart, both > 22 KHz, at several different frequencies. IOW one test would be composed of tones at 24 and 28 KHz, another test composed of signals at 32 and 36 KHz, and another at 38 and 42 KHz. Be careful with these test tones, as it would be easy to turn them up high enough to involve large amounts of power without you hearing that they are loud. If there is excess nonlinear distoriton in your system you'll hear a 4 KHz tone.

Another, potentially safer test signal would be white or pink noise high pass filtered at 22 KHz. If there is excess nonlinear distoriton in your system you'll hear noise in the normal audio band.

This post has been edited by db1989: Oct 19 2011, 09:58

Reason for edit: holy moly why quote that massive first post in its entirety

[googlebot]

Wouldn't IM artifacts within the audible band have been audible with the Cantons also? Sadly I had to let the Elacs go, because their WAF was too low. I'm looking for a satellite/sub combo with comparable bandwidth right now but haven't decided on anything, yet. The test sample has very high amounts of HF energy, do you have any experience with IM distortions in plain air? Or would that be unrealistic under the given test conditions (especially the ~48kHz lowpass).

This post has been edited by db1989: Oct 19 2011, 09:59

Reason for edit: deleting unneeded quote of above post

[Arnold B. Kruege...]

Wouldn't IM artifacts within the audible band have been audible with the Cantons also?

The IM could be in the speakers, or a consequence of amp/speaker interfacing. IM in speakers is not uncommon. Other tests of high sample rates have been invalidated by nonlinearities in the speakers and amplifiers.

Sadly I had to let the Elacs go, because their WAF was too low. I'm looking for a satellite/sub combo with comparable bandwidth right now but haven't decided on anything, yet.
The test sample has very high amounts of HF energy, do you have any experience with IM distortions in plain air?

In general, air is far more linear than speakers.

This post has been edited by db1989: Oct 19 2011, 09:59

Reason for edit: truncating quote

[JSW]

I just tried an ABX with this sample. To eliminate any effects of HF noise, I made two 16/96 files from the original 24/192, both with TPDF flat dither, one of them via floating-point/44.1. At first I thought I could hear a difference between them, but the test results were typical: 6 out of 14 correct. I still think that maybe I could hear the difference at first.

BTW, until recently I was unequipped to perform this kind of test. The Java ABX/HR program choked on 96kHz files (on my computer at least) and rejected 24-bit ones. I recently got a new(er) computer running OSX 10.5 and just loaded ABXer. I'll be performing several more tests in the next few days.

[JSW]

I just did a listening test on the effect of downsampling a 96/24 file to 44.1, and the results were not random. I used the same piece of music as I used in last fall's
dithering test,
which is track number 8 of

http://www.hdtracks.com/index.php?file=cat...=HD030911109622


I took the commercially available 96/24 file, converted it to 44.1/fp with Audacity ("high quality sinc filter") and then back to 96/24 (sample rate conversion with Audacity, floating-point to 24bit with ffmpeg). I ABX'ed using the Mac program ABXer, which for reasons I have detailed elsewhere is a PITA, and through my good loudspeakers. If you want the commercially available file and you're outside the United States, reportedly you can buy it by using Paypal but not by using a credit card with non-US billing address. By listening for a part of the sound that was highly dependent on the angle I held my head, I concluded that "B" was "File 2", the ultrahigh bandwidth version. Here is the transcript:

ABX Test Completed: 2010-07-03 13:47:12 -0500

Number of tests performed: 12
Number of correct answers: 0
Percentage correct: 0%

File 1 = /Users/jeffrey/Music/littlered_lowpass_24_96.wav
File 2 = /Users/jeffrey/Music/iTunes/iTunes Media/Music/Unknown Artist/Symphonic Dances (hi-rez)/8-EtudesTableaux Little Red Ri.aiff
File placement was static.

n [A] [X] [B] Choice Score
1 [2] [1] [1] A 0/1
2 [2] [2] [1] B 0/2
3 [2] [2] [1] B 0/3
4 [2] [1] [1] A 0/4
5 [2] [2] [1] B 0/5
6 [2] [1] [1] A 0/6
7 [2] [2] [1] B 0/7
8 [2] [1] [1] A 0/8
9 [2] [1] [1] A 0/9
10 [2] [1] [1] A 0/10
11 [2] [1] [1] A 0/11
12 [2] [1] [1] A 0/12

--------------------------------------------------------------

I strongly suspect that this result reflects a bug in the program. The Preferences dialog in ABXer has a checkbox "randomize files," which I do not have checked. Most likely this means that it is following my preference to map File 1 -> A, File 2 -> B, but it interpreted my answers using the other mapping.

[Notat]

You might want to exclude nonlinear distortion in the playback system as a potential source of problems with the test by doing some listening tests with test files composed of sine waves 4 KHz apart, both > 22 KHz, at several different frequencies. IOW one test would be composed of tones at 24 and 28 KHz, another test composed of signals at 32 and 36 KHz, and another at 38 and 42 KHz. Be careful with these test tones, as it would be easy to turn them up high enough to involve large amounts of power without you hearing that they are loud. If there is excess nonlinear distoriton in your system you'll hear a 4 KHz tone.

Another, potentially safer test signal would be white or pink noise high pass filtered at 22 KHz. If there is excess nonlinear distoriton in your system you'll hear noise in the normal audio band.

What is your threshold for "excess nonlinear distortion"? Are there speakers or headphones that can reproduce ultrasonics and pass these tests?

[Arnold B. Kruege...]

I can definitately hear 0.1 % nonlinear distortion in an ABX test if I get to pick the program material. Below that, things get hard and I make no claims about hearing 0.03% under any even vaguely real world circumstances.

There are defintely loudspeaker systems that can reproduce >20 KHz, at least if you sit close and on-axis. I built a pair of them using some Danish ring radiator tweeters. They seemed capable enough of producing random guessing in tests involving the audibility of ultrassonic signals that they are probably free of problematical nonlinear distortion.

This post has been edited by db1989: Oct 19 2011, 10:00

Reason for edit: removing pointless quote of above post

[JSW]

I don't think I'm prepared to perform either of these tests. I'd need a level meter to interpret any 4kHz tone, and as for the broadband noise, I'm not sure it would be diagnostic, since I can already hear the idle channel noise of the amplifier. What would a slight change in its color mean?

I have a Rotel RA-1062 driving a pair of B&W 704's. The Rotel is supposed to have a good enough damping ratio to control any potential ultrasonic resonances in the tweeters. When I bought my system, I didn't know what technical issues might come up, but I figured I was paying premium prices for something engineered well beyond any reasonable standards and could stop worrying about this kind of thing.

This post has been edited by db1989: Oct 19 2011, 10:00

Reason for edit: ditto

[onlyconnect]

I am surprised this thread did not prompt more discussion. Is this evidence for the effectiveness of hypersonics? Is this proof that 24/96 is audibly different than 16/44 on the right equipment?

Tim

[Arnold B. Kruege...]

Certain 24/96 recordings will sound audibly different than 16/44 if the monitoring chain contains certain technical defects that are not uncommon.

That's old news that has been replayed with a certain amount of fanfare during the past decade.

This post has been edited by db1989: Oct 19 2011, 10:00

Reason for edit: as above

[IgorC]

Here ABX test was successful again on those samples 16/44.1 vs 24/96. Now it's completely different hardware. It can't be coincidence.

[Canar]

24/96 vs. 16/44.1 has two independent variables. To properly isolate variables, it's best to compare 24/44.1 with 16/44.1, 16/96 with 16/44.1, and 24/96 with 24/44.1.

[IgorC]

Good point, Canar,

I was thinking too about trying it such way.

I can see that there are two groups of people with extreme beliefs.
One group think that is absolutely nothing above 16/44.1 and other group (mostly audiofools) believe that the difference is more than obvious between 16/44.1 and 24/96.

There is also 3d group of people (me too) who think that the truth is somewhere in the middle (but still thinks that in 99% of cases 16/44.1 is enough). This group believes that there is a very small % of people who can really hear more than that and the difference isn't easy to spot. I can ABX (again and again, with different hardware and conditions) 16/44.1 vs 24/96. But it's actually enough subtle and not so obvious as lossy compression at high bitrates or 32 kHz vs 44.1 kHz.

This post has been edited by IgorC: Oct 19 2011, 04:52

[onlyconnect]

I can ABX (again and again, with different hardware and conditions) 16/44.1 vs 24/96. But it's actually enough subtle and not so obvious as lossy compression at high bitrates or 32 kHz vs 44.1 kHz.

I find your position plausible in general but have had less ABX success than you. What equipment do you use? Are you successful with both loudspeakers and headphones? Are these samples you have made yourself, and how are you downsampling?

Tim