foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/16 19:01:08
File A: C:\Users\KiarkAudio\Documents\Ravel_Test_File_2.wav
File B: C:\Users\KiarkAudio\Documents\Ravel_Test_File_1.mp3
19:01:08 : Test started.
19:01:27 : 01/01 50.0%
19:01:34 : 02/02 25.0%
19:01:42 : 03/03 12.5%
19:02:15 : 04/04 6.3%
19:02:49 : 05/05 3.1%
19:03:01 : 06/06 1.6%
19:03:52 : 07/07 0.8%
19:04:06 : 08/08 0.4%
19:04:28 : 09/09 0.2%
19:04:38 : 10/10 0.1%
19:04:51 : 11/11 0.0%
19:05:28 : 12/12 0.0%
19:05:44 : Test finished.
----------
Total: 12/12 (0.0%)
This test is from an excerpt of Ravel's Daphnis et Chloe (full version, not the Suite). This was a planned set of 12 rounds.
The .wav file above was ripped directly from the music CD using SoundForge 10. SoundForge 10 exported the file to MP3 at 320 kbs setting.
Soundforge statistics produced the following:
Redbook Left RMS: -26.087
Redbook Right RMS: -26.385
MP3 Left RMS: -26.135
MP3 Right RMS: -26.441
While the two files are thus not identically matched for level, the Left channel difference is 0.048db, Right difference is 0.056 db. While these are not
perfectly matched, I don't claim to hear that small a difference, and particularly not when averaged across the file. Minimum and Maximum samples were all slightly over or under -6db, I assume the recording was mastered to -6db. Thus, not near clipping level.
The test above was done listening to the opening portion of the clip.
Here's a <30-second clip of the lossless source:
http://www.hydrogenaudio.org/forums/index....ost&id=7640 (http://www.hydrogenaudio.org/forums/index.php?act=attach&type=post&id=7640)
Here is the thread containing it:
http://www.hydrogenaudio.org/forums/index....howtopic=102681 (http://www.hydrogenaudio.org/forums/index.php?showtopic=102681)
Playback chain is:
Lenovo PC desktop, files saved to hard drive
RME Babyface audio interface
Schiit Asgard 2 headphone amplifier
Beyerdynamic DT 770 Pro headphones
I'm posting this because I've read in other forums (not here; just joined) that MP3 at 256 kbs and 320 kbs are "transparent", indistinguishable from Redbook audio. I chose 320 as the highest quality available in MP3 - the difference is audible to me. They are certainly very close - listening in my car driving down the road I wouldn't hear the difference.
I have lots of professional musical experience, but my 50+ year-old ears apparently do not hear above 18 kHz. I know the basic principles of psychoacoustic masking used in MP3 lossy compression, but not well enough to know whether or not Ravel in live orchestra is more suited or less suited to accurate MP3 encoding. Alternative sample files for testing welcome.
Not all mp3 encoders are created equally?
Have you tried Lame 3.98.4, 3.99.5 or 3.100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322)?
This last one uses VBR which could likely do just as well as 320 CBR (which is true for VBR with the other two versions).
Also, don't worry about matching level unless the codec scales as a precaution to prevent clipping on decode to integer. That you get different values is perfectly normal with lossy encoding.
Depends. ABXIng some codecs at 320kbps isn't very hard. Is the one you used any good?
Just in case your program did something funky:
Could you convert the "wav file" to 320 kbps (-b 320) using the lame command line. I suggest to use the most default setting possible.
While you are it try VBR 0 (-V 0 ).
Not all mp3 encoders are created equally?
Have you tried Lame 3.98.4, 3.99.5 or 3.100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322)?
This last one uses VBR which could likely do just as well as 320 CBR (which is true for VBR with the other two versions).
Also, don't worry about matching level unless the codec scales as a precaution to prevent clipping on decode to integer. That you get different values is perfectly normal with lossy encoding.
Not yet. I have SoundForge, so I used it. Just downloaded 3.99.5 from RareWares, I'll have to figure out the command line. Thanks for the tip.
Just downloaded 3.99.5 from RareWares, I'll have to figure out the command line. Thanks for the tip.
Try with, xrecode it's easy, no need to "type" command lines. The software download lame encoder for you, it's almost automated. Xrecode have a nag screen, but is fully functional if you use the "demo".
Xrecode have a nag screen, but is fully functional if you use the "demo".
Instead you could use foobar2000 or LameDropXPd.
The test above was done listening to the opening portion of the clip.
I'm interested in evaluating the killer sample you've found when you can upload another version.
(Insert long-winded justification of the statement that either the OP has golden ears or there's something wrong here)
Not saying that the OP did anything wrong, just that this sort of result shouldn't happen very often to people who don't have golden ears, and when it does it can be considered a bug.
Just in case your program did something funky:
Could you convert the "wav file" to 320 kbps (-b 320) using the lame command line. I suggest to use the most default setting possible.
While you are it try VBR 0 (-V 0 ).
Okay, here we are.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/16 22:15:57
File A: C:\Users\KiarkAudio\Documents\Ravel_Test_File_2.wav
File B: C:\Users\KiarkAudio\Documents\Ravel_Test_File_3.mp3
22:15:57 : Test started.
22:16:13 : 01/01 50.0%
22:16:22 : 02/02 25.0%
22:16:30 : 03/03 12.5%
22:16:40 : 04/04 6.3%
22:16:49 : 05/05 3.1%
22:17:08 : 06/06 1.6%
22:17:24 : 07/07 0.8%
22:18:15 : 08/08 0.4%
22:20:03 : 09/09 0.2%
22:21:17 : 10/10 0.1%
22:21:50 : 11/11 0.0%
22:22:28 : 11/12 0.3%
22:22:48 : Test finished.
----------
Total: 11/12 (0.3%)
Test File 3 was made with LAME 3.99.5 vc6.
Command line switches were -b 320 -h
A planned set of 12 rounds again.
I'd like to get around the "killer sample" limitation. That's the reason for using this clip: wide variety of orchestral textures, good dynamic range while not pushing the clipping limits. I hope to replicate from different portions of the clip. This was listening *only* to the opening few bars.
^^ Unless that music is in the public domain you're going to get yourself in trouble posting links like that. Anyway, without the original lossless file, its hard to gauge anything about that mp3. You need both to compare.
Edit: You may want to look at TOS#9 now: http://www.hydrogenaudio.org/forums/index....974#entry149482 (http://www.hydrogenaudio.org/forums/index.php?showtopic=3974#entry149482)
Sorry about the length mistake, moderators, it won't happen again!
Proper versions well under 30 seconds:
Original lossless version, WAV ripped from Redbook CD:
http://www.hydrogenaudio.org/forums/index....ost&id=7640 (http://www.hydrogenaudio.org/forums/index.php?act=attach&type=post&id=7640) (same as the link in the discussion's first post)
Recommended MP3 via LAME 3.99.5, at -b 320 -h settings:
http://www.hydrogenaudio.org/forums/index....ost&id=7641 (http://www.hydrogenaudio.org/forums/index.php?act=attach&type=post&id=7641)
I can't hear any difference between the two. I quit after 25 minutes, clicking a guess so as to obtain a log.
foo_abx 1.3.4 report
foobar2000 v1.2.9
2013/09/16 21:20:02
File A: C:\Users\Admin\Desktop\Ravel_Test_File_2_short.wav
File B: C:\Users\Admin\Desktop\Ravel_Test_File_3_short.mp3
21:20:02 : Test started.
21:45:08 : 00/01 100.0%
21:45:13 : Test finished.
----------
Total: 0/1 (100.0%)
I don't have the ears (or the equipment?) to ABX that high
Can you also try the VBR 0 (eg: -V 0), just a curiosity.
EDIT:
I'm under the impression that the internals are different enough to merit different results.
I don't have the ears (or the equipment?) to ABX that high
Can you also try the VBR 0 (eg: -V 0), just a curiosity.
EDIT:
I'm under the impression that the internals are different enough to merit different results.
Help me keep up with the terminology here
What are the internals? Are you referring to the different processing employed by the -V 0 switch, compared to the -b 320 setting for the posted MP3?
Thanks for the help! Will definitely try out the -V 0 switch tomorrow.
Also, I don't want to depend on this one clip--if someone has already prepared clips of this specific contrast (44.1/16 versus high-quality MP3), I would like to try them out.
Maybe we got ourselves a new Killer Sample? It's certainly possible...
Anyway, I tried ABXing. I failed, unsurprisingly...
Recommended MP3 via LAME, at -b 320 -h
Is -h actually recommended? For some reason my brain is telling me it's not, but I'll admit to not being totally up on all the LAME switches.
http://wiki.hydrogenaudio.org/index.php?ti...ncoder_settings (http://wiki.hydrogenaudio.org/index.php?title=LAME#Recommended_encoder_settings)
Recommended MP3 via LAME, at -b 320 -h
Is -h actually recommended? For some reason my brain is telling me it's not, but I'll admit to not being totally up on all the LAME switches.
The RareWares bundle includes a "Basic command line switch reference" HTML file, reproduced here:
http://ecmc.rochester.edu/ecmc/docs/lame3.98a2/basic.html#h (http://ecmc.rochester.edu/ecmc/docs/lame3.98a2/basic.html#h)
...which recommends -h for slower encoding but higher quality. I'm interested in distinguishing at the highest possible quality levels, so I used it.
I believe there are documented cases where -h resulted in a poorer quality encode
Anyway, the 30-second lossless clip is adequate; people can encode it themselves.
You're free to upload to our forum rather than use a shitty third-party ad-laden host.
Equipment may or may not make a difference, depending on the artifact. More-expensive does not universally mean more revealing. Grossly uneven frequency response might be able to break masking, while excessive distortion or noise can aid masking.
Lastly, MP3 has limitations which hinder its ability to deliver transparent results for those who are trained to listen for them or who are particularly sensitive to them.
http://lame.cvs.sourceforge.net/viewvc/lam...l/detailed.html (http://lame.cvs.sourceforge.net/viewvc/lame/lame/doc/html/detailed.html)
Aliases and removed settings
-h Alias of -q 2.
[...]
-q n Algorithm quality selection
For CBR, ABR and --vbr-old modes, the following table applies
-q 0 Use slowest & best possible version of all algorithms.
-q 3 Default value. Good speed, good quality
[...]
For the default VBR mode since LAME 3.98, the following table applies
-q 0 to -q 4 Use the best algorithm.
@UltimateMusicSnob:
If you like to try lame3100l mentioned by greynol I suggest to use the setting --bCVBR 316. It uses a very high quality VBR method internally (which needs an audio data bitrate like CBR 320).
I believe there are documented cases where -h resulted in a poorer quality encode
This must be what I'm remembering; the rare cases that have come up where it was worse. It's not necessarily a bad or frowned upon switch, then.
...
File A: C:\Users\KiarkAudio\Documents\Ravel_Test_File_2.wav
File B: C:\Users\KiarkAudio\Documents\Ravel_Test_File_3.mp3
...
I'd like to get around the "killer sample" limitation...
Why is it your mp3 has a different name as the wav?
Since you are new in here i have some doubts about your intention and motivation, sorry.
You may have only abx'd some hamster fart against the zar bomb.
One reason could be that windows does hide file extension by default, and if you have the same media icon for all media files, that makes it even worse.
So you would have a:
Ravel_Test_File_1 with a Music Note Icon, and
Ravel_Test_File_1 with a Music Note Icon
But one is actually a WAV and the other an MP3, or something else, etc.
Not everybody changes the windows default and turns extensions back on.
...
File A: C:\Users\KiarkAudio\Documents\Ravel_Test_File_2.wav
File B: C:\Users\KiarkAudio\Documents\Ravel_Test_File_3.mp3
...
I'd like to get around the "killer sample" limitation...
Why is it your mp3 has a different name as the wav?
Since you are new in here i have some doubts about your intention and motivation, sorry.
You may have only abx'd some hamster fart against the zar bomb.
When I do these sorts of tests, I usually end up with multiple files whose only difference is the precise encoding employed. In this case one lossless source file, used for multiple lossy versions in different encodings.
So I number them sequentially as they are produced, so that they line up properly when I sort on the 'Name' column in Windows Explorer.
If I don't distinguish them by number, then successive varieties of MP3 will over-write each other.
The hamster business, well, that could be true of anyone putting up foobar2000 logs. For that matter, I could have just edited the file in Notepad to say whatever I like. Short of doing the tests in a face-to-face environment, I don't see a way around that one.
Finally, "intention and motivation" are immaterial to objective testing. Folks in Gearslutz said that Hydrogen Audio was highly knowledgeable about objective testing for audio, so that's my interest.
FWIW, I think the thinly-veiled accusation was out of line. The vast majority here are willing to take your results at face value until there is reason to think otherwise.
Lastly, MP3 has limitations which hinder its ability to deliver transparent results for those who are trained to listen for them or who are particularly sensitive to them.
This is a difference I'm interested in. On other forums, discussions of whether the much greater space required by Redbook quality digital files is "worth it" have produced occasional claims that "no one can hear the difference between Redbook and high-quality MP3 anyway", while I would have said "most people", or "casual listeners", or "listeners in noisy ambient environments", etc.
I am not trained to listen specifically for the limitations of MP3, but I am trained to listen to classical music at excruciating levels of detail (B.M., M.M., D.M.A., Music Composition). I suspect that drives both my interest, and potentially to the possibility of results as well.
Ultimate, could you please upload the wav in the Uploads forum? This zippyshare website is b0rken and should probably be nuked from orbit anyway.Oh, you already have. Okay then.
Since you are new in here i have some doubts about your intention and motivation, sorry.
That's why we ask people to provide samples. Peer review and all that.
Ultimate, could you please upload the wav in the Uploads forum? This zippyshare website is b0rken and should probably be nuked from orbit anyway.
Oh, you already have. Okay then.
Since you are new in here i have some doubts about your intention and motivation, sorry.
That's why we ask people to provide samples. Peer review and all that.
I'd love to see replication. One outlier, unless replicated, will always just be one outlier, without significance in the larger scientific understanding of the phenomena in question.
As far as I can tell, the MP3 encoding algorithms are in great shape. You gotta be kind of obsessed, just to figure out how to listen for any potential limitations of the lossy versions, let alone detect them reliably. I'm definitely obsessed.
Definitely agree on zippyshare--I was just defaulting to a practice employed elsewhere, on forums which did not provide Upload capacity. Sorry about that, the practice here is far superior, kudos to the owners/designers.
I can't even begin to discern a difference, even after I removed the sample padding from the mp3 and boosted both by 20dB (!).
Maybe you can tell us what you heard differently?
It's nothing to do with high frquencies, since there's no HF to speak of.
...
File A: C:\Users\KiarkAudio\Documents\Ravel_Test_File_2.wav
File B: C:\Users\KiarkAudio\Documents\Ravel_Test_File_3.mp3
...
I'd like to get around the "killer sample" limitation...
Why is it your mp3 has a different name as the wav?
Since you are new in here i have some doubts about your intention and motivation, sorry.
You may have only abx'd some hamster fart against the zar bomb.
If you have any
specific doubts, it might be helpful to raise them.
If not, and you just have a "bad feeling", then you're bringing no more to the discussion is the kind of unscientific, biased thinking that ABX tests were designed to circumvent.
I can't even begin to discern a difference, even after I removed the sample padding from the mp3 and boosted both by 20dB (!).
Maybe you can tell us what you heard differently?
It's nothing to do with high frquencies, since there's no HF to speak of.
After the fact I went to SPAN Voxengo plug-in in Sound Forge, and as you say there's no HF to speak of. In that respect it's a poor sample--I was thinking orchestration and musical texture when I picked this one, not frequency spectra per se.
If I were setting up a guitar tone, I would call the MP3 version more 'scooped' - a little more bass, and little more high (of what little there is), and the WAV version more balanced. It's subtle, but once I locked onto that as a characteristic, I went through pretty quickly. The incidental sounds of the concert hall in the very very beginning pop a little more in the MP3 version, seems to me.
You may want to try 3.100 alpha 2 (http://www.rarewares.org/mp3-lame-bundle.php#lame-alpha), or 3100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322) (which is a modification of the former).
There was a recent case (http://www.hydrogenaudio.org/forums/index.php?showtopic=100256) of a member who could ABX 3.99.5 @ 320 kbps, but couldn't ABX 3.100 alpha 2, AAC 320 and AAC vbr 200.
You may want to try 3.100 alpha 2 (http://www.rarewares.org/mp3-lame-bundle.php#lame-alpha), or 3100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322) (which is a modification of the former).
There was a recent case (http://www.hydrogenaudio.org/forums/index.php?showtopic=100256) of a member who could ABX 3.99.5 @ 320 kbps, but couldn't ABX 3.100 alpha 2, AAC 320 and AAC vbr 200.
Thanks for the tips, I will try those. I have been reading about the quality of AAC going back at least to the early 2000's, I think.
Anyone here remember LiquidAudio???
I was able to successfully ABX the Lame 3.99.5 sample what was prepared by UltimateMusicSnob as well as one encoded using my preferred mp3 profile of Lame 3.98.4 -V3. Lame 3.98.4 -b320 gave me trouble, but I might have a better idea what to listen for after spending a lot of time with the sample encoded with 3.99.5.
Results for 3.99.5 -h -b320:
foo_abx 1.3.4 report
foobar2000 v1.2.9
2013/09/17 15:35:57
File A: C:\Ravel_Test_File_2_short.flac
File B: C:\Ravel_Test_File_3_short.mp3
15:35:57 : Test started.
15:43:51 : 01/01 50.0%
15:50:17 : 02/02 25.0%
15:50:47 : 03/03 12.5%
15:51:02 : 04/04 6.3%
15:51:27 : 05/05 3.1%
15:51:46 : 06/06 1.6%
15:52:23 : 06/07 6.3%
15:52:53 : 07/08 3.5%
15:53:04 : 08/09 2.0%
15:53:31 : 08/10 5.5%
15:53:57 : 09/11 3.3%
15:54:14 : 10/12 1.9%
15:54:24 : Test finished.
----------
Total: 10/12 (1.9%)
I was able to successfully ABX the Lame 3.99.5 sample what was prepared by UltimateMusicSnob as well as one encoded using my preferred mp3 profile of Lame 3.98.4 -V3. Lame 3.98.4 -b320 gave me trouble, but I might have a better idea what to listen for after spending a lot of time with the sample encoded with 3.99.5.
Results for 3.99.5 -h -b320:
foo_abx 1.3.4 report
foobar2000 v1.2.9
2013/09/17 15:35:57
File A: C:\Ravel_Test_File_2_short.flac
File B: C:\Ravel_Test_File_3_short.mp3
15:35:57 : Test started.
15:43:51 : 01/01 50.0%
15:50:17 : 02/02 25.0%
15:50:47 : 03/03 12.5%
15:51:02 : 04/04 6.3%
15:51:27 : 05/05 3.1%
15:51:46 : 06/06 1.6%
15:52:23 : 06/07 6.3%
15:52:53 : 07/08 3.5%
15:53:04 : 08/09 2.0%
15:53:31 : 08/10 5.5%
15:53:57 : 09/11 3.3%
15:54:14 : 10/12 1.9%
15:54:24 : Test finished.
----------
Total: 10/12 (1.9%)
Replication! Do you think the low signal levels in this clip contributed to the ability to get a positive result? WAY back in the late 90's, I MP3-ed some live classical recordings which required quite low signals for some passages, and they broke up entirely under the algorithm I had back then (whatever SoundForge 4 had back then.
Replication! Do you think the low signal levels in this clip contributed to the ability to get a positive result? WAY back in the late 90's, I MP3-ed some live classical recordings which required quite low signals for some passages, and they broke up entirely under the algorithm I had back then (whatever SoundForge 4 had back then.
Pretty sure it does. When on my PC i listen this i had to put the level on maximum to hear it at somewhow normal level. I still don't know what to listen for
I remember for vbr tuning there once had a compromise to be found. Wasting much bits on silent parts made no sense to some and when the bitrate increased to satisfy reproducing these silent parts there were loud complaints about bloating bitrate.
For high bitrate cbr i can guess there never went to much tuning into these situations because there are more other obvious problems to solve.
Edit': for info, i use an X-Fi driving HD-590 at the PC, it works pretty nicely but has its limitations.
I don't have the ears (or the equipment?) to ABX that high
Can you also try the VBR 0 (eg: -V 0), just a curiosity.
EDIT:
I'm under the impression that the internals are different enough to merit different results.
Okay, here's the foobar result for the same source clip encoded in LAME 3.99.5 using the VBR switch set at -V0. Planned set of 12.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/17 21:33:35
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_4.mp3
21:33:35 : Test started.
21:34:23 : 01/01 50.0%
21:36:14 : 02/02 25.0%
21:36:31 : 03/03 12.5%
21:37:10 : 04/04 6.3%
21:37:39 : 05/05 3.1%
21:38:24 : 06/06 1.6%
21:39:48 : 06/07 6.3%
21:40:48 : 07/08 3.5%
21:41:57 : 08/09 2.0%
21:43:02 : 09/10 1.1%
21:44:01 : 10/11 0.6%
21:46:34 : 10/12 1.9%
21:46:46 : Test finished.
----------
Total: 10/12 (1.9%)
Did not behave precisely the same as the -b 320 using CBR. Seemed better behaved on the high frequencies, still had the difference in bass that I described above as contributing to "slightly scooped". Another difference here is that I went to the middle-latter portion of the track for a stronger signal, while the previous tests were done from the beginning. VERY subtle differences of course.
I noticed when messing with the 'lead voice' sample that I was able to abx it simply because the mp3 version was subtly louder than the source. Is this a common thing for LAME to do or is it just the peculiarity of that sample?
Also I noticed the OP is using Beyerdynamic headphones, which are known to be rather bright and highly detailed to the point of being described as fatiguing. I've been able to abx (http://www.hydrogenaudio.org/forums/index.php?showtopic=96329) high bit rate mp3 by making adjustments to the upper mid/treble eq bands while using much more mundane iems.
Of course I am NOT questioning the OP's position, merely pointing out to any doubters that it's not impossible to abx high bitrate mp3s, regardless of 50 year old ears. I hear better at 35 than 15 because of learning/training. Age has nothing to do with it.
@greynol: And you claimed to not have 'golden ears'...
I noticed when messing with the 'lead voice' sample that I was able to abx it simply because the mp3 version was subtly louder than the source. Is this a common thing for LAME to do or is it just the peculiarity of that sample?
Also I noticed the OP is using Beyerdynamic headphones, which are known to be rather bright and highly detailed to the point of being described as fatiguing. I've been able to abx (http://www.hydrogenaudio.org/forums/index.php?showtopic=96329) high bit rate mp3 by making adjustments to the upper mid/treble eq bands while using much more mundane iems.
Of course I am NOT questioning the OP's position, merely pointing out to any doubters that it's not impossible to abx high bitrate mp3s, regardless of 50 year old ears. I hear better at 35 than 15 because of learning/training. Age has nothing to do with it.
@greynol: And you claimed to not have 'golden ears'...
Thanks for the thoughtful comments. Apologies for not knowing the terminology, what is the "lead voice" sample? I did check levels on the very first pair and saw differences in the 4 or 5 hundredths of a decibel range, which I figured was not detectable.
I'm surprised that Beyerdynamics are considered bright--my DT 770 Pro's strike me as darker than everything else I've ever listened on. My old (REALLY old) DT 48's, yeah, those are bright.
Of course I didn't EQ anything, but it makes sense to test that way, since minimally knowledgeable users
will apply EQ in their listening for enjoyment, even if all they have is Bass and Treble controls.
Very glad to hear independent confirms that MP3 is not perfectly transparent (contra folks in a couple of other audio forums, not here).
Lead voice is one of the 'killer samples' that have been known to give encoders fits even at high bitrates, it's been used to fine tune LAME in attempts at transparency. You can get it here http://www.hydrogenaudio.org/forums/index....showtopic=50056 (http://www.hydrogenaudio.org/forums/index.php?showtopic=50056)
I could be mistaken about your particular model of headphones, I don't have any experience other than reviews about Beyerdynamics in general being 'bright', but on second look some agree with your dark assertion. Grain of salt I guess.
... Do you think the low signal levels in this clip contributed to the ability to get a positive result?...
If this is the case lame3100l --bCVBR 316 may well improve the situation.
As a rule of thumb, if you are going to apply eq then apply it before encoding. The encoder naturally assumes that no eq will be subsequently applied and will make decisions based on that assumption.
As a rule of thumb, if you are going to apply eq then apply it before encoding. The encoder naturally assumes that no eq will be subsequently applied and will make decisions based on that assumption.
I know that makes sense from a purely psychoacoustic point of view, but few people do this. You'd have to re-encode everything each time you wanted to change the EQ!
Always-on EQ is typically used to compensate for specific speaker or headphones, or match personal taste. I'm not about to re-encode my entire collection every time I change my headphones, and I don't listen to my collection on just
one set of headphones / speakers anyway.
Stating the obvious here Just don't want the OP to think that everyone here encodes mp3s with EQ baked in. I'm 100% sure they don't.
Cheers,
David.
compensate for specific speaker or headphones
If done properly, this will not break masking.
compensate for specific speaker or headphones
If done properly, this will not break masking.
True. With poor speakers+headphones,
not using EQ "could" break masking.
Cheers,
David.
Right. I suspect that uneven headphones are responsible for good number of positive ABX results.
I could be mistaken about your particular model of headphones, I don't have any experience other than reviews about Beyerdynamics in general being 'bright', but on second look some agree with your dark assertion. Grain of salt I guess.
It is bright in terms of having peaky treble in the 5-10 kHz octave, but it also has a notch in the upper midrange at about 3-4 kHz (this (http://en.goldenears.net/index.php?mid=GR_Headphones&category=275&page=3&document_srl=8535) frequency response graph does not match exactly what I hear, but it does show those issues). So, masking in that range could indeed be affected.
Right. I suspect that uneven headphones are responsible for good number of positive ABX results.
Does it rise to the level of becoming something that has to (or at least, potentially could) be addressed in the MP3 algorithms themselves?
What I'm thinking of is the millions of listeners who *do* use MP3's nearly all the time, on cheap earbuds, cheap earphones, cheap desktop speakers. That's really the target market for MP3 files: portable, small, non-audiophile playback chains.
Mix and mastering engineers of course are already well aware that their products will be played back on sub-optimal systems, and have dedicated techniques and 'bad' reference monitors just for that purpose, to see how robust the mix is in different contexts (nothing is robust to cheap OR expensive earbuds, IMO, but oh well...).
It is bright in terms of having peaky treble in the 5-10 kHz octave, but it also has a notch in the upper midrange at about 3-4 kHz (this (http://en.goldenears.net/index.php?mid=GR_Headphones&category=275&page=3&document_srl=8535) frequency response graph does not match exactly what I hear, but it does show those issues). So, masking in that range could indeed be affected.
So that we're clear, this will not turn into a discussion about headphones.
Does it rise to the level of becoming something that has to (or at least, potentially could) be addressed in the MP3 algorithms themselves?
I don't think so. Those interested in transparency will use lossless or perhaps a more modern lossy format. Once you are at the level of near-transparency, ignoring killer samples/hard to encode instruments (harpsichord), due to broken masking or otherwise, the differences are so subtle that they'll likely never be noticed except during a critical listening test where you have the lossless version quickly accessible.
Mix and mastering engineers of course are already well aware that their products will be played back on sub-optimal systems, and have dedicated techniques and 'bad' reference monitors just for that purpose, to see how robust the mix is in different contexts (nothing is robust to cheap OR expensive earbuds, IMO, but oh well...).
I don't have much faith in sound "engineers" when it comes to their ability to assess objective sound quality to the point that I must question their assessment of subjective sound quality. Perhaps, for me, it has been spoiled by just a few nut-jobs; oh well.
To drag this back toward 'Listening Tests', I mostly do my foobar trials with closed-back headphones on. It's a purely circumstantial choice, as I have children running around the house, not particularly quiet A/C switching on and off, plenty of ambient and incidental noise generally. Without isolation, I doubt I could concentrate, let alone test.
There are probably some differences in outcomes due to testing with headphones w/no crossfeed instead of speakers. I'm assuming speakers would be more difficult, given that all other factors (including some commensurable measure of "accuracy" or "precision" in the headphone or speaker quality) were the same, but perhaps that's not the case.
In my noisy environment, it's either use the headphones or don't bother. Even with that, a good number of my tests are of the type, Run 5 rounds, fix two bowls of cereal, do 5 more.
In my noisy environment, it's either use the headphones or don't bother. Even with that, a good number of my tests are of the type, Run 5 rounds, fix two bowls of cereal, do 5 more.
This methodology, far from being suboptimal, probably also keeps you from a significant amount of fatigue.
It's often recommended around these parts to take breaks during listening tests. You've stumbled upon a best practice by accident!
You've stumbled upon a best practice by accident!
I don't think it was solely by accident. It seems UltimateMusicSnob has quiete a bit of experience and tried it the correct way from posting results over posting samples to being still polite while being offended very unpolitely by me..
My excuse for that but lately there come more and more pinheads to forums only to claim all kinds of extraordinary things they don't take in question. Leave alone prove something with a valid testing process.
My excuse for that but lately there come more and more pinheads to forums only to claim all kinds of extraordinary things they don't take in question. Leave alone prove something with a valid testing process.
I agree it's nice to help someone who tries to do things right. Some people just can't be helped (I wonder many times why they bother posting...) but it's certainly not the case here, where the OP is showing not only to be willing to do what we kindly tell him to, but also a very knowledgeable person. Kudos to him.
On topic,
I mostly do my foobar trials with closed-back headphones on
It's generally asserted that you can detect encoding artifacts easier by using headphones, even more if their frequency response is not even, as was previously said. IRC most developers of lossy codecs fine-tune their algorithms with headphones.
Even so, at 320kbps you should not be hearing a difference either way. Interesting case we got here...
There are probably some differences in outcomes due to testing with headphones w/no crossfeed instead of speakers.
There may be but not necessarily so, it depends on which set of speakers and which headphones were used in that test and how true to the source each of them are.
Even so, at 320kbps you should not be hearing a difference either way. Interesting case we got here...
Let's not go that far with the absolutes (bolding mine). It's quite rare to hear a difference at such bitrates, but it does happen.
If this is the case lame3100l --bCVBR 316 may well improve the situation.
halb's lame3100 has been mentioned several times in this thread. Do you plan on trying it out, UltimateMusicSnob?
Even so, at 320kbps you should not be hearing a difference either way. Interesting case we got here...
Let's not go that far with the absolutes (bolding mine). It's quite rare to hear a difference at such bitrates, but it does happen.
If this is the case lame3100l --bCVBR 316 may well improve the situation.
halb's lame3100 has been mentioned several times in this thread. Do you plan on trying it out, UltimateMusicSnob?
Absolutely, yes. I've done SF 10, LAME at 320 and at -V0. Frankly, each one requires listening for different characteristics
to discern a difference. 3100 is next.
Even so, at 320kbps you should not be hearing a difference either way. Interesting case we got here...
Let's not go that far with the absolutes (bolding mine). It's quite rare to hear a difference at such bitrates, but it does happen.
Yes, you're right. My apologies.
at 320kbps you should not be hearing a difference either way
I took you to mean: "I think that 320kbps should be sufficient to achieve transparency"? (And when it fails it is likely due to a shortcoming of the encoder/codec.)
at 320kbps you should not be hearing a difference either way
I took you to mean: "I think that 320kbps should be sufficient to achieve transparency"? (And when it fails it is likely due to a shortcoming of the encoder/codec.)
I meant that, yes, but I agree that the phrase
should not is a bit extreme and I was not very clear. Never mind.
EDIT: Likely due to a shortcoming of the encoder, but possibly because of the person listening and his hearing.
There are people who can routinely discern 320 from lossless on non-killer samples.
There are people who can routinely discern 320 from lossless on non-killer samples.
And I know that. Poor people...
Okay, this was LAME 3100 at V0. This version reminded me most of the Sound Forge encoding I used in the first example, although the bass weighting was slightly different (less as compared to SF10).
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/18 21:40:27
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_5.mp3
21:40:27 : Test started.
21:40:54 : 01/01 50.0%
21:41:12 : 02/02 25.0%
21:41:41 : 03/03 12.5%
21:42:03 : 04/04 6.3%
21:42:28 : 05/05 3.1%
21:44:26 : 06/06 1.6%
21:45:26 : 07/07 0.8%
21:52:23 : 08/08 0.4%
21:53:16 : 08/09 2.0%
21:54:01 : 08/10 5.5%
21:54:18 : 09/11 3.3%
21:54:45 : 10/12 1.9%
21:54:58 : Test finished.
----------
Total: 10/12 (1.9%)
There are people who can routinely discern 320 from lossless on non-killer samples.
Another question on nomenclature: a "killer sample" is one which has some feature that produces a relatively obvious artifact? The comments I've seen about "lead sample" here and there give me this impression.
The distinction I'm wondering about is between a sample that some can positive-ABX just because they hear a quality difference of *some* kind, and a second sample (possibly the "killer") in which some artifact pops out because it contains some sonic characteristic which causes something more like a breakdown of the effectiveness of the MP3 algorithm being used?
The Ravel, for example, is very low level, but low levels can be found in thousands of recordings--I wouldn't think of this as a killer sample, then. Or am I missing the meaning altogether?
Okay, this was LAME 3100 at V0. This version reminded me most of the Sound Forge encoding I used in the first example, although the bass weighting was slightly different (less as compared to SF10).
I just gave ABXing this clip a shot. At -V5 I think I was on to something, but after 5 trials or so I seemed to lose it. Maybe if I'd taken a break I'd have done better. Regardless, I think I can safely say ABXing this clip at any sort of high bitrate is out of my league.
I don't meant to pester you into another test, but with halb's lame3100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322), perhaps you should try the settings he (http://www.hydrogenaudio.org/forums/index.php?showtopic=102679&view=findpost&p=845069) himself (http://www.hydrogenaudio.org/forums/index.php?showtopic=102679&view=findpost&p=845163) recommended earlier in this thread: --bCVBR 316
Okay, this was LAME 3100 at V0. This version reminded me most of the Sound Forge encoding I used in the first example, although the bass weighting was slightly different (less as compared to SF10).
I just gave ABXing this clip a shot. At -V5 I think I was on to something, but after 5 trials or so I seemed to lose it. Maybe if I'd taken a break I'd have done better. Regardless, I think I can safely say ABXing this clip at any sort of high bitrate is out of my league.
I don't meant to pester you into another test, but with halb's lame3100l (http://www.hydrogenaudio.org/forums/index.php?showtopic=102322), perhaps you should try the settings he (http://www.hydrogenaudio.org/forums/index.php?showtopic=102679&view=findpost&p=845069) himself (http://www.hydrogenaudio.org/forums/index.php?showtopic=102679&view=findpost&p=845163) recommended earlier in this thread: --bCVBR 316
Thanks for the reminder. I'll do that setting next.
Thank you for testing Lame 3100 (Lame 3.100a2?, lame3100l?) -V0. It's really remarkable that you got at 5/5 within just two minutes.
But if the issue is really due to low volume, -V0 is hardly the solution, no matter used with Lame 3.100a2 or lame3100l. I'm looking forward to your --bCVBR 316 result.
Thank you for testing Lame 3100 (Lame 3.100a2?, lame3100l?) -V0. It's really remarkable that you got at 5/5 within just two minutes.
But if the issue is really due to low volume, -V0 is hardly the solution, no matter used with Lame 3.100a2 or lame3100l. I'm looking forward to your --bCVBR 316 result.
That was LAME 3100L. Actually 3100l, but then it's hard to make out what the last character is--I had to hunt around in Command Prompt to find the directory ZIP just created, thought I had a capital I there.
For the target market and usage of MP3 generally, I'm coming to the opinion that the sort of ABX-ing I'm doing is mostly beside the point.
I hunt for audible differences in order to get these results. They are subtle, and they do take some hunting. Once I find the audible distinctions, completing the ABX with positive result
can go quickly. Not always--depends on what I'm listening for--bass can take longer than treble, for example.
Aside from demonstrating that MP3 is not *perfectly* transparent for *all* listeners under excruciatingly detailed inspection--which is in fact my research question--it's difficult for me to see how these results could advance the state of the art at all. All I've got at the end is a 'black swan', per John Stuart Mill, falsifying a universal proposition.
IMO your listening test is valuable, especially as it's about 'normal' music, not special electronic stuff or harpsichord or other specials where it is known that mp3 can behave way below perfection.
It's also valuable for people to get a personal attitude towards your results and what they mean to them that you're adding comments like 'the difference is subtle'.
Thanks again.
BTW I also make the difference between listening tests and practical listening situations. I'm paranoid enough not to accept issues which could be audible in practical listening situations even when chance for that is very low. But there are issues that are audible to me in ABX tests when comparing carefully against the original (the harp40_1 sample for instance) but which are so subtle that to me they can be ignored for practical listening situations.
especially as it's about 'normal' music
...albeit poorly recorded 'normal' music.
at 320kbps you should not be hearing a difference either way
I took you to mean: "I think that 320kbps should be sufficient to achieve transparency"? (And when it fails it is likely due to a shortcoming of the encoder/codec.)
I would say it's a shortcoming of a whole MP3 format. Just try electronic music with sharp transients. There was sample of electronic music, "EIG". It will never be transparent for MP3 320 kbps. Any MP3 encoder will have some pre-echo on such samples because the short block is still quite big (192) comparing to other formats (commonly 120 or 128). That's why AAC/Opus/Vobis do better job at ~192 kbps than MP3 at 320 kbps (considering high quality encoders for all formats).
P.S. Link to eig sample http://www.hydrogenaudio.org/forums/index....showtopic=49601 (http://www.hydrogenaudio.org/forums/index.php?showtopic=49601)
a "killer sample" is one which has some feature that produces a relatively obvious artifact?
Yes, they are samples that expose an encoders weaknesses; though they aren't necessarily so obvious that everyone can identify the artifacts.
The comments I've seen about "lead sample" here and there give me this impression.
I have no idea what a "lead sample" is.
(possibly the "killer") in which some artifact pops out because it contains some sonic characteristic which causes something more like a breakdown of the effectiveness of the MP3 algorithm being used?
Correct.
The Ravel, for example, is very low level, but low levels can be found in thousands of recordings--I wouldn't think of this as a killer sample, then. Or am I missing the meaning altogether?
I honestly don't know if this can be categorized as a killer sample. Is the entire recording this low, or does it reach full-scale (or within a couple of dB or full-scalle)? Someone coughing next to the microphone doesn't count. Honestly, this is a horrible recording. I'm not at all surprised that a lossy encoder is having a hard time making a determination of what constitutes signal.
In the end I think it's a matter of whether it's getting the noise right. I don't agree with the idea that there is a difference in tonal balance.
.. There was sample of electronic music, "EIG". ..
While you're totally right that mp3 isn't perfect at electronic stuff like this I think mp3 can be very acceptable even with this. eig is one of those killer samples that made me create lame3100l, and in this case it doesn't take highest bitrate settings to make it very acceptable IMO. Did you try say lame3100l -V0 on eig?
.. There was sample of electronic music, "EIG". ..
While you're totally right that mp3 isn't perfect at electronic stuff like this I think mp3 can be very acceptable even with this. eig is one of those killer samples that made me create lame3100l, and in this case it doesn't take highest bitrate settings to make it very acceptable IMO. Did you try say lame3100l -V0 on eig?
habl27,
It was a several years. 3.97, 3.98, 3.99, all your extensions 3.99/3.100x..... A bit better handling but still the issue is here.
All these years LAME devs have made all possible to make it better. And it's got better but it's still there.
You can do an extraodrinady job but You can't overcome a limitation of format. No.
especially as it's about 'normal' music
...albeit poorly recorded 'normal' music.
A live orchestral performance in the hall, captured to a recording, no matter how good the mic's or the engineer, always sounds 'canned' and 'distant' to me, and no amount of extra mic's or later EQ-ing has ever fixed that in my experience. A full Romantic orchestra can peak 120db or even higher (http://www.soundadvice.info/thewholestory/san12.htm) (a lone operatic soprano can do nearly as well), so the dynamic range is extraordinary, and classical musicians are loath to give *anything* up to compression.
This clip starts out extraordinarily low, but in the climaxes later on it reaches to its maximum mastered level of slightly over -6db (-5.7xx db, if I recall). Is there some other sense in which this clip is badly recorded?
I also think the recording is okay. The spot is just low volume. Not so rare with Ravel: when testing Musepack recently it was the quiet starting part of Ravel's Bolero where I was able to ABX Musepack.
And with a longer part that low it's also quite common to turn up volume, because at home most people can't realize the dynamics of a concert hall.
So I think it's an interesting relevant sample.
And sure I'm curious about lame3100l --bCVBR 316's quality which should take good care of low volume spots.
This clip starts out extraordinarily low, but in the climaxes later on it reaches to its maximum mastered level of slightly over -6db (-5.7xx db, if I recall).
To be clear, and not to insinuate that you do, but turning the volume up to hear differences in detail for the quiet parts constitutes cheating in my book, unless you normally ride the volume while listening.
To me this is no different than cranking the gain on fade-outs and reverb tails in order to proclaim that 16 bits are insufficient as a delivery format.
Is there some other sense in which this clip is badly recorded?
I don't find any other issues with the quality of the sample you provided. That there's no annoying hum actually suggests to me the recording is pretty good, now that I know what kind of dynamic range it has. It could make better use of 16 bits considering that the peak just makes use of the most significant bit. This is assuming it was originally recorded as >16-bit digital. If it's from analog tape then this last part about using more than 16 bits is not relevant.
This clip starts out extraordinarily low, but in the climaxes later on it reaches to its maximum mastered level of slightly over -6db (-5.7xx db, if I recall).
To be clear, and not to insinuate that you do, but turning the volume up to hear differences in detail for the quiet parts constitutes cheating in my book, unless you normally ride the volume while listening.
To me this is no different than cranking the gain on fade-outs and reverb tails in order to proclaim that 16 bits are insufficient as a delivery format.
Is there some other sense in which this clip is badly recorded?
I don't find any other issues with the quality of the sample you provided. That there's no annoying hum actually suggests to me the recording is pretty good, now that I know what kind of dynamic range it has. It could make better use of 16 bits considering that the peak just makes use of the most significant bit. This is assuming it was originally recorded as >16-bit digital. If it's from analog tape then this last part about using more than 16 bits is not relevant.
The label says recorded in digital, and it's early (80s), so I would guess original capture was 16 bits, but I can't confirm directly.
I ride the volume all the time in the car, at home I don't ride much, but I do tend to run pretty loud overall, classical/rock/techno alike, max ~92 db measured at my nearfield position. Headphones have better isolation to start with so don't need to be so loud; I'm also paranoid about misjudging with headphones on, so I probably max a little less in the loudest parts, 88db or so.
I would say testing protocol specifics should be determined by the research question of interest. To improve the algorithm itself, as a developer I would care most about conditions typical of listeners--that's why I posted earlier that the tests posted here are of limited application. I have done cheap-earphones ABX tests before--might be interesting to repeat some trials that way.
I ride the volume all the time in the car
I don't advise conducting ABX testing in the car under typical conditions, though I have a feeling you won't pass.
Not that you didn't already admit that.
Okay, this is a replication under different playback hardware.
These are <$10 in-ear earphones. Apologies, but there's no brand name listed on them, so I don't know the make. Overall, I'd call them "dull" by comparison to my Beyerdynamics, and maybe weak in bass (no surprise), a little harsh on mids, not too bad.
Very good seal actually, so isolation is at least close to comparable.
Powered by on-the-motherboard audio chip for Dell Optiplex 780. Manual says it's "ADI 1984A High Definition Audio Codec - Integrated on System Board" ("Codec"???). Device Manager is even less forthcoming, just calls it "High Definition Audio Device".
This is with File 3, which was the originally requested LAME 3.99.5_vc6 encoder.
Planned set of 12.
I have another requested encoding to do for 3100L, otherwise I think there's nothing more to be learned from this Ravel, unless someone has a point they want to pursue.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/19 16:50:18
File A: E:\Ravel_Test_File_2_short.wav
File B: E:\Ravel_Test_File_3_short.mp3
16:50:18 : Test started.
16:50:48 : 01/01 50.0%
16:51:18 : 02/02 25.0%
16:51:28 : 02/03 50.0%
16:51:44 : 03/04 31.3%
16:52:00 : 04/05 18.8%
16:52:49 : 05/06 10.9%
16:53:09 : 06/07 6.3%
16:53:30 : 07/08 3.5%
16:54:30 : 08/09 2.0%
16:55:22 : 09/10 1.1%
16:56:52 : 10/11 0.6%
17:00:00 : 10/12 1.9%
17:00:08 : Test finished.
----------
Total: 10/12 (1.9%)
One thing to mention related to some people always hear some defects with mp3 coding can have a simple hearing defect as reason. When mp3 was pretty new one of the first public listening test by c't magazine germany had one clear winner. The man was well trained but also had a hearing defect that made some filtering not working correctly om him.
Of course digging in the noisefloor for errors makes not much sense also like greynol already mentioned.
One thing to mention related to some people always hear some defects with mp3 coding can have a simple hearing defect as reason. When mp3 was pretty new one of the first public listening test by c't magazine germany had one clear winner. The man was well trained but also had a hearing defect that made some filtering not working correctly om him.
Of course digging in the noisefloor for errors makes not much sense also like greynol already mentioned.
Interesting angle. Got a citation?
German only, maybe some onlinetranslator will help.
http://www.heise.de/ct/artikel/Kreuzverhoertest-287592.html (http://www.heise.de/ct/artikel/Kreuzverhoertest-287592.html)
The comments I've seen about "lead sample" here and there give me this impression.
I have no idea what a "lead sample" is.
He's referring to lead voice (http://www.hydrogenaudio.org/forums/index.php?act=attach&type=post&id=2704), which I referred to earlier, and is perhaps the most annoying thing I've ever abx tested with.
Okay, Ravel encoded in LAME 3100L using the recommended --bCVBR 316.
Planned set of 12 rounds.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/20 00:36:46
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_6.mp3
00:36:46 : Test started.
00:37:01 : 01/01 50.0%
00:37:13 : 02/02 25.0%
00:37:24 : 03/03 12.5%
00:37:49 : 03/04 31.3%
00:38:10 : 04/05 18.8%
00:38:29 : 05/06 10.9%
00:38:47 : 06/07 6.3%
00:39:16 : 07/08 3.5%
00:39:44 : 07/09 9.0%
00:40:22 : 08/10 5.5%
00:41:08 : 09/11 3.3%
00:43:59 : 10/12 1.9%
00:44:02 : Test finished.
----------
Total: 10/12 (1.9%)
I deliberately set out
not to hunt for specific artifacts this time, but to decide on the basis of a gestalt. My cue here was "focus", meaning more for Redbook and a more diffuse sound for MP3. I would expect a similar sort of difference if I took an existing track and chorused and reverbed it just slightly. As an artifact, it's rather a pleasant one in terms of the aura of sound; the downside would be slightly less precision in the definition of separate instruments.
Of the encodings I've tried, this last one matches my taste preferences best. It seems to have frequency bands best balanced, among the MP3's.
German only, maybe some onlinetranslator will help.
http://www.heise.de/ct/artikel/Kreuzverhoertest-287592.html (http://www.heise.de/ct/artikel/Kreuzverhoertest-287592.html)
Thanks for the cite, very interesting. Turns out my German (former life as a dramatic tenor--Florestan, Bacchus, Siegmund) is still pretty good. If anyone's interested, the article reports a listening test subject with hearing damaged by an "explosion accident" which left him with a usable hearing range only up to 8 kHz, and tinnitus. So "bright" noises which in normal listeners would trigger the masking effects don't register with him, leaving the flanging artifacts of MP3 exposed. If I were studying psychoacoustics professionally, I would want to find out what this listener's responses reveal about audio processing---potentially very useful, in the same way that persons suffering from brain lesions in various regions shed valuable light on how the brain works.
Last time I checked I still hear up to about 18 kHz, which represents the hearing loss more or less normal to aging.
Okay, Ravel encoded in LAME 3100L using the recommended --bCVBR 316.
Planned set of 12 rounds.
Thanks for posting. I was curious how 3100l would do. Looks like halb has some explaining to do. Or, perhaps, tuning.
Thanks for the cite, very interesting. Turns out my German (former life as a dramatic tenor--Florestan, Bacchus, Siegmund) is still pretty good. If anyone's interested, the article reports a listening test subject with hearing damaged by an "explosion accident" which left him with a usable hearing range only up to 8 kHz, and tinnitus. So "bright" noises which in normal listeners would trigger the masking effects don't register with him, leaving the flanging artifacts of MP3 exposed. If I were studying psychoacoustics professionally, I would want to find out what this listener's responses reveal about audio processing---potentially very useful, in the same way that persons suffering from brain lesions in various regions shed valuable light on how the brain works.
Last time I checked I still hear up to about 18 kHz, which represents the hearing loss more or less normal to aging.
I guess several people that have dips at one or more frequency bands will also be able to hear the masking not working properly. In the factory my father worked many people had a dip at some special frequency from the machines there running 24/7.
I guess several people that have dips at one or more frequency bands will also be able to hear the masking not working properly. In the factory my father worked many people had a dip at some special frequency from the machines there running 24/7.
Conceivably, but in that case it would depend on what frequency bands, of what width, and with what degree of hearing loss, interacting with what program material. The researcher would want to ask listeners with those hearing characteristics to listen for flanging effects. But unless someone knows of completed research in this area already, it's an unconfirmed hypothesis.
Even with his extreme hearing loss, the man in the article had a hit rate of only 90%, and that was comparing 128 kbs MP3 (from an encoder written no later than March of 2000) to CD. I would ordinarily consider that a failure in ABX, since I use 95% as the threshold for significance, and prefer to go way beyond that.
Okay, Ravel encoded in LAME 3100L using the recommended --bCVBR 316.
Planned set of 12 rounds.
Thanks for posting. I was curious how 3100l would do. Looks like halb has some explaining to do. Or, perhaps, tuning.
I intended to give him at least some reason for celebrating. It's a very effective algorithm, and it does handle the low levels well, I thought.
VBR sees low levels and thinks "I don't need to code this stuff, I can save some bits by not encoding it". CBR sees low levels and thinks "well, I've got to use these bits somewhere so I might as well encode this stuff". Difference is huge.
I'm oversimplifying. e.g. decent VBR will dynamically adjust absolute thresholds depending on signal level, because it knows one use case is turning the volume up to make low level stuff perfectly audible.
I don't know if this issue is just down to levels though. It would be trivial to boost the sample (at least the quiet part) before encoding to check.
EDIT: You've got to be very careful with such low levels.
ReplayGain = +17.81dB
Track peak = 0.077026
A simple truncation to 15-bits is audible...
foo_abx 1.3.3 report
foobar2000 v1.2.2
2013/09/20 15:08:07
File A: D:\audio\audio\codec testers\Ravel_Test_File_2_short.wav
File B: D:\audio\audio\codec testers\Ravel_Test_File_2_short 15-bit trunc.wav
15:08:07 : Test started.
15:08:59 : 01/01 50.0%
15:09:03 : 02/02 25.0%
15:09:10 : 03/03 12.5%
15:09:17 : 04/04 6.3%
15:09:24 : 05/05 3.1%
15:09:32 : 05/06 10.9%
15:09:43 : 06/07 6.3%
15:10:07 : 07/08 3.5%
15:10:10 : Test finished.
----------
Total: 7/8 (3.5%)
(dithering to 15-bits is even more audible)
EDIT2:
Oh blimey, even re-dithering to 16-bit is audible...
foo_abx 1.3.3 report
foobar2000 v1.2.2
2013/09/20 15:15:43
File A: D:\audio\audio\codec testers\Ravel_Test_File_2_short.wav
File B: D:\audio\audio\codec testers\Ravel_Test_File_2_short 16bit dither.wav
15:15:43 : Test started.
15:16:13 : 01/01 50.0%
15:16:22 : 02/02 25.0%
15:16:32 : 03/03 12.5%
15:16:38 : 03/04 31.3%
15:16:45 : 04/05 18.8%
15:16:50 : 05/06 10.9%
15:16:54 : 06/07 6.3%
15:17:05 : 07/08 3.5%
15:17:07 : Test finished.
----------
Total: 7/8 (3.5%)
Unless you're using a proper 24-bit mp3 decoder with 24-bit output, just the decoding could be enough to audibly alter this file if you listen at a very raised volume appropriate to make this very quiet part easily audible.
Boosting it by exactly 800% before encoding will at least remove this issue. I assume somewhere there's a peak above -18dB, so you can't actually do this for the whole track - but then I bet you can't listen at such a loud level for the whole track either.
ABXing quiet sections with the volume cranked up is kind of extreme.
Cheers,
David.
Okay, Ravel encoded in LAME 3100L using the recommended --bCVBR 316. ...(result: 7/8 after only 2.5 minutes)
...Of the encodings I've tried, this last one matches my taste preferences best. It seems to have frequency bands best balanced, among the MP3's.
Thanks for testing.
a) I'm happy to see that lame3100l --bCVBR 316 improves the situation a little bit and seems to be superior to CBR320.
b) I'm disappointed to see that ABXing seemed to be pretty easy for you also with this encoder and setting.
I've encoded the sample myself and looked whether there's something wrong with 3100l. Seems not to be the case: average bitrate is 318 kbps for the mp3packed file, so the constrained VBR method makes full use of the maximum data space possible also for this low volume snippet. Looking at the individual frames everything is as is expected, too: it's all long blocks with the exception of one frame at the very end, and they're all encoded with highest bitrate. Increase of SNR compared to plain -V0 is about 9 db for every frame.
Open question (if I haven't overlooked something): did you crank up the volume when ABXing compared to your usual listening level?
[pointless full quote of giant post removed]
On the tests above, what were you listening for? Is this listening for a noise floor? If noise floor is just white noise, then the only thing to listen for is levels of noise, yes? A difference in how much white noise you hear below the music, when the volume is cranked way up? And of course, if not for the noise floor, then what is the audible cue?
Okay, Ravel encoded in LAME 3100L using the recommended --bCVBR 316. ...(result: 7/8 after only 2.5 minutes)
...Of the encodings I've tried, this last one matches my taste preferences best. It seems to have frequency bands best balanced, among the MP3's.
Thanks for testing.
a) I'm happy to see that lame3100l --bCVBR 316 improves the situation a little bit and seems to be superior to CBR320.
b) I'm disappinted to see that ABXing seemed to be pretty easy for you also with this encoder and setting.
I've encoded the sample myself and looked whether there's something wrong with 3100l. Seems not to be the case: average bitrate is 318 kbps for the mp3packed file, so the constrained VBR method takes full use of the maximum data space possible also for this low volume snippet. Looking at the individual frames everything is as is expected, too: it's all long blocks with the exception of one frame at the very end, and they're all encoded with highest bitrate. Increase of SNR compared to plain -V0 is about 9 db for every frame.
Open question (if I haven't overlooked something): did you crank up the volume when ABXing compared to your usual listening level?
No, did not crank. I need to caution persons looking at the results--this was not easy. It requires
very close mental focus, and a clear and well-defined notion of what to listen for. In this case I wanted to test non-artifact impressions strictly, so the notion was "focus". Turning up the volume makes that more difficult, not easier. This was hard, and the hardest part is keeping your mental ears' attention on the characteristics you're looking for. Especially because ABX trials are repetitions of the same clip, the mind tends to get, if not bored outright, then say "over-accustomed" to the sounds--they become too familiar, so attention wanders.
This is a great encoder, as I say. My research question is "Is it possible at all?", and my conclusion would be at this point, Yes, it's possible, but you have to be kind of obsessive about close listening.
On the tests above, what were you listening for? Is this listening for a noise floor? If noise floor is just white noise, then the only thing to listen for is levels of noise, yes? A difference in how much white noise you hear below the music, when the volume is cranked way up? And of course, if not for the noise floor, then what is the audible cue?
Yes, noise floor. The noise floor of the recording isn't quite white. With dither there's perfectly white noise added on top. Without dither (truncation) the added noise is quieter but fractionally less uniform. All fairly subtle though and very hard to detect without a reference.
Volume was cranked up, but I would happily listen to the full 30 second clip at that volume - it's not loud. I wouldn't dream of listening to anything else at anywhere near that volume though!
It's interesting that lossyWAV doesn't even touch the first 5 seconds of the file (it maintains the full 16-bits unchanged), and throughout the rest of the file it never knocks off more than 1-bit from the right channel.
Cheers,
David.
Something this low in volume should compress extremely well, even without lossyWAV.
Assuming we're still talking about lossy, what makes you say that?
Sorry, I should have made myself more clear.
2Bdecided commented on how lossyWAV needed to keep nearly every bit due to the low signal level. I was just pointing out that as far as lossless compression is concerned, you don't need lossyWAV anyway because the compression to lossless will be so efficient.
Sorry if this was too off topic.
Oh, I get it now.
I think David often likes to use lossyWAV as a tool to determine how many LSBs can be tossed without affecting audibility, based on a setting that hasn't yet been shown not to be transparent?
I might have suggested that the OP try to ABX the lossyWAV version, but clearly that would be a waste of time in this case.
I'm wondering whether this sample will have an influence on further mp3 development.
That halb27's fork does so well is encouraging.
I also wonder how well AAC handles this, though not enough to put myself through another ABX session.
I'm wondering whether this sample will have an influence on further mp3 development.
This seems to be a real corner case, and even the OP doesn't seem to feel that it needs to be fixed or improved in any way. Personally I would treat it as a curious anomaly and leave it at that.
What I am much more interested in is the OP's ability to hear the subtlest of details. He could possibly be a valuable asset in assessing codec differences.
On the tests above, what were you listening for? Is this listening for a noise floor? If noise floor is just white noise, then the only thing to listen for is levels of noise, yes? A difference in how much white noise you hear below the music, when the volume is cranked way up? And of course, if not for the noise floor, then what is the audible cue?
Yes, noise floor. The noise floor of the recording isn't quite white. With dither there's perfectly white noise added on top. Without dither (truncation) the added noise is quieter but fractionally less uniform. All fairly subtle though and very hard to detect without a reference.
Volume was cranked up, but I would happily listen to the full 30 second clip at that volume - it's not loud. I wouldn't dream of listening to anything else at anywhere near that volume though!
It's interesting that lossyWAV doesn't even touch the first 5 seconds of the file (it maintains the full 16-bits unchanged), and throughout the rest of the file it never knocks off more than 1-bit from the right channel.
Cheers,
David.
This is something that it appears will be a fundamental difference between classical and pop repertoire now indefinitely, thanks to the inherent nature of the material, but also the loudness wars. One reason good stereo equipment for classical becomes expensive is that the signal is 'only' normalized to -6db to begin with, plus you have these very soft passages. So the amplification stages had better be very quiet indeed, with plenty of headroom. On speakers, I prefer to listen to classical on 3 or 4 out of 10 on the dial, but on a big powerful amp stage with big speakers.
And then of course the pop music now normalizes to, well, often 0 db, and if not 0 then -0.1 db.
The consequence is that no single MP3 algorithm could ever be optimized for these dramatically disparate groups of program material. As a business proposition I'd figure all the hard work for MP3 development goes into pop music, since the difference in sales figures is what, 2 or 3 orders of magnitude?
I'm wondering whether this sample will have an influence on further mp3 development.
This seems to be a real corner case, and even the OP doesn't seem to feel that it needs to be fixed or improved in any way. Personally I would treat it as a curious anomaly and leave it at that.
What I am much more interested in is the OP's ability to hear the subtlest of details. He could possibly be a valuable asset in assessing codec differences.
Yes, it's definitely edge of the envelope. greynol did replicate, but that just means two outliers instead of one on the record for this clip.
As I said at the end of the OP, Alternative sample files are **welcome**. These research questions are of great interest to me (http://books.google.com/books?id=5fLgxlcDv3wC&printsec=frontcover&dq=jerald+hughes+transmutation+of+digital+information+goods&hl=en&sa=X&ei=YJQ8UvOMAajc2QWep4HYAg&ved=0CC8Q6AEwAA#v=onepage&q=jerald%20hughes%20transmutation%20of%20digital%20information%20goods&f=false). Also, details of protocol matter, as the research questions vary, so finding out about improvements to testing methodology is also of great interest.
I think David often likes to use lossyWAV as a tool to determine how many LSBs can be tossed without affecting audibility, based on a setting that hasn't yet been shown not to be transparent?
Yes, exactly.
It's also a helpful indicator of where the 20Hz-15kHz noise floor lies - which in my mind is another way of saying the same thing. Others may disagree. It might not be true that you can always quantise at x dB below a noise floor calculated by some reasonable method and maintain transparency - but it seems to work well enough.
Cheers,
David.
These research questions are of great interest to me (http://books.google.com/books?id=5fLgxlcDv3wC&printsec=frontcover&dq=jerald+hughes+transmutation+of+digital+information+goods&hl=en&sa=X&ei=YJQ8UvOMAajc2QWep4HYAg&ved=0CC8Q6AEwAA#v=onepage&q=jerald%20hughes%20transmutation%20of%20digital%20information%20goods&f=false).
Is that you, or just a link showing what you're interested in?
These research questions are of great interest to me (http://books.google.com/books?id=5fLgxlcDv3wC&printsec=frontcover&dq=jerald+hughes+transmutation+of+digital+information+goods&hl=en&sa=X&ei=YJQ8UvOMAajc2QWep4HYAg&ved=0CC8Q6AEwAA#v=onepage&q=jerald%20hughes%20transmutation%20of%20digital%20information%20goods&f=false).
Is that you, or just a link showing what you're interested in?
That's me.
... Personally I would treat it as a curious anomaly ...
An anomaly?
We're used to considering issues in stuff like electronic music, harpsichord music, or other relatively special contents as an anomaly (as long as we are not very much into these genres), and we ignore them more or less, or use very high quality mp3 which brings even these issues pretty close to transparency.
But this is a sample without any special a priori characteristics for being an issue to mp3. And the issue has been confirmed.
The good news is that this can be considered not relevant in practical listening situations, in case I understand UltimateMusicSnob correctly (please tell us if I'm wrong).
Well, I think in the end it's nothing that essentially changes our attitude towards the usage of mp3. We have known before that there are spots in the music where we can get only near-transparency. What's new is that this doesn't happen only to special stuff, but also to pretty 'normal' music.
An interesting question is:
Claiming the issue being irrelevant is (hopefully) correct for people who use very high quality mp3.
But what about the people who use say -V2, a setting which is considered to yield a very good quality/filesize relation?
UltimateMusicSnob, would you mind trying -V2 and tell us what you think about quality in practical listening situations?
(sorry I can't contribute myself with listening here - I cannot hear any deviation from the original even with lower quality settings than -V2).
Since that short sample has a headroom of more as 22dB it may be worth to shift it losslessly 18dB louder so no dither can influence the noise. UMS may test if the encode suffers by its pure loudness or by other things when brought back to higher levels.
Edit: Anyone noticed the left channel only being 15bit in that sample from he beginning? No wonder lossywav can't throw away much. Very srange to have it only on one channel.
... Personally I would treat it as a curious anomaly ...
An anomaly?
We're used to considering issues in stuff like electronic music, harpsichord music, or other relatively special contents as an anomaly (as long as we are not very much into these genres), and we ignore them more or less, or use very high quality mp3 which brings even these issues pretty close to transparency.
But this is a sample without any special a priori characteristics for being an issue to mp3. And the issue has been confirmed.
The good news is that this can be considered not relevant in practical listening situations, in case I understand UltimateMusicSnob correctly (please tell us if I'm wrong). UMS: This is correct,
Well, I think in the end it's nothing that essentially changes our attitude towards the usage of mp3. We have known before that there are spots in the music where we can get only near-transparency. What's new is that this doesn't happen only to special stuff, but also to pretty 'normal' music.
An interesting question is:
Claiming the issue being irrelevant is (hopefully) correct for people who use very high quality mp3.
But what about the people who use say -V2, a setting which is considered to yield a very good quality/filesize relation?
UltimateMusicSnob, would you mind trying -V2 and tell us what you think about quality in practical listening situations?
(sorry I can't contribute myself with listening here - I cannot hear any deviation from the original even with lower quality settings than -V2).
I would happy to. Unless you say otherwise I'll assume that -V2 is the only switch for the command line.
Yes, just -V2.
Okay, Ravel clip with 3100L -V2. All planned sets of 12 rounds each.
Before the results, general comments:
Very good algorithm, holds up extremely well in comparison to the -V0. All differences reported here are also subtle. Shares the mid-scooped characteristics of all MP3 encodings, still very subtle here. Shares the slightly diffused bass of -V0, but the 'aura' imparted to the sound is still very musical, an aurally pleasing result. Trade-off appears to be accuracy in preserving exact timbre of the instruments' sounds, for the 'bloom' or 'aura' that results. Transients?
I chose three listening targets ahead of time, all non-artifact.
First target was 'quality of bass'. Listened at 3.0 to about 9 seconds for the crescendo.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/20 21:05:44
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_6.mp3
21:05:44 : Test started.
21:06:37 : 01/01 50.0%
21:06:53 : 02/02 25.0%
21:07:34 : 03/03 12.5%
21:08:04 : 03/04 31.3%
21:08:38 : 04/05 18.8%
21:09:12 : 05/06 10.9%
21:10:01 : 06/07 6.3%
21:10:32 : 07/08 3.5%
21:11:37 : 08/09 2.0%
21:12:19 : 09/10 1.1%
21:12:55 : 10/11 0.6%
21:14:30 : 10/12 1.9%
21:14:35 : Test finished.
----------
Total: 10/12 (1.9%)
Cues used:
WAV bass has smooth, focused crescendo.
MP3 bass crescendo seems to proceed by plateaus, diffuse.
Second target was 'quality of treble'. Listened at 0 seconds to about 6 seconds.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/21 09:22:12
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_6.mp3
09:22:12 : Test started.
09:22:36 : 00/01 100.0%
09:22:46 : 01/02 75.0%
09:23:02 : 02/03 50.0%
09:23:14 : 03/04 31.3%
09:23:29 : 04/05 18.8%
09:23:54 : 05/06 10.9%
09:24:15 : 06/07 6.3%
09:25:18 : 07/08 3.5%
09:27:23 : 08/09 2.0%
09:28:27 : 09/10 1.1%
09:29:55 : 10/11 0.6%
09:31:14 : 10/12 1.9%
09:31:25 : Test finished.
----------
Total: 10/12 (1.9%)
Cues used:
Treble is just a bit brighter in MP3, plus some room sound seems to be missing. Can't hear the same reverb reflections in the MP3.
Third target was "flutes" as an impression of the instrument sound, rather than a frequency band. Concentrated on the sforzando flute attack at 8.0 seconds and following.
foo_abx 1.3.4 report
foobar2000 v1.2.8
2013/09/21 11:25:15
File A: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_2_short.wav
File B: C:\Users\KiarkAudio\Documents\Ravel Listening Tests\Ravel_Test_File_6.mp3
11:25:15 : Test started.
11:25:38 : 01/01 50.0%
11:25:52 : 01/02 75.0%
11:26:08 : 02/03 50.0%
11:26:23 : 03/04 31.3%
11:26:36 : 04/05 18.8%
11:28:11 : 05/06 10.9%
11:30:40 : 06/07 6.3%
11:32:42 : 07/08 3.5%
11:33:24 : 08/09 2.0%
11:35:39 : 09/10 1.1%
11:37:26 : 10/11 0.6%
11:39:20 : 10/12 1.9%
11:39:22 : Test finished.
----------
Total: 10/12 (1.9%)
Cues used:
MP3 Flute attack at 8.0 seconds is less crisp than WAV. Flute timbre a [u]little[/u] blurred, compared to focused sound of WAV.
Bass is diffuse, like the other trials.
Thanks a lot for your test.
So it looks like -V2 quality is also okay for this sample in practical listening situations.
I'd like to add that other than with --bCVBR 316 special version lame3100l has no influence on this result. I've looked it up, and because of the low volume lame3100l does not increase SNR (with the exception of the last two frames which contain short resp. mixed blocks). As was to be expected. So Lame 3.100a2 should work just as fine.
Thanks a lot for your test.
So it looks like -V2 quality is also okay for this sample in practical listening situations.
I'd like to add that other than with --bCVBR 316 special version lame3100l has no influence on this result. I've looked it up, and because of the low volume lame3100l does not increase SNR (with the exception of the last two frames which contain short resp. mixed blocks). As was to be expected. So Lame 3.100a2 should work just as fine.
Do you already know where on the -V scale the outcome re SNR will begin to differ? That inflection point is probably something to know about.
Also, I was just thinking about the psychoacoustic phenomenon of masking--for these very soft passages, what masking could possibly be occurring? Little or none?--so there may not be much for the core theoretical idea of MP3 to grab hold of here.
I've tried -V2 to -V0 here: lame3100l's SNR increasing feature doesn't come into play here. An SNR increase of 3..4 db starts with settings like --bCVBR 266.
The reason behind this is that standard Lame behaves so well. Average bitrate of Lame3.100a2 for your sample and -V2/-V1/-V0 is 173/201/226 kbps despite being low volume. It's 195/228/264 for my standard test set of various pop music which is rather loud overall compared to your sample. lame3100l does a lot for the quality of short and mixed blocks even for moderate quality settings. But this is not relevant here. At -V2 or better it also takes care of long blocks which are relevant here. It's done in a way which doesn't increase average bitrate considerably. (For a stronger bitrate increase using a more demanding -V level is expected to be more appropriate). lame3100l's constrained VBR mechanism works by not allowing audio data bitrate to go too low (not below 153 kbps in the case of -V2 resp. 226 kbps in the case of -V0).
This minimum audio data bitrate is energy dependent however, and the values given are used for a reference level of moderate loudness (which actually is ~10 times the energy level of your sample - so not very much higher; high volume pop music has an energy ~100 times higher than my reference level). So what essentially keeps lame3100l -Vx from improving the situation is that standard Lame does so well in the first place, and also by the way I do the energy dependent SNR increase. As for possible changes concerning the last point see my next post, please.
When there's little masking a good VBR method should take this into account and behave accordingly. Looks like Lame (standard, not just my variant) does a good job at this.
I'm wondering whether this sample will have an influence on further mp3 development.
As far as the development of my Lame variant is concerned the sample has an influence.
This sample shows that the way I do constrained VBR depending on energy level is pretty questionable.
So I guess that I'll create another version where CVBR will have only a very weak energy dependence relevant only for extremely low volume spots.
I do not expect this to significantly improve lame3100l's quality for this sample however, at least not when using -Vx. For using -Vx I will still take care that average bitrate increase will be negligible compared to standard Lame. Whenever standard Lame works adequately as with this sample I do not expect this modification will come into play with -Vx (maybe a bit when using -V0).
Improvement may come with settings like --bCVBR 275, but whether that's audible is another question.
Since that short sample has a headroom of more as 22dB it may be worth to shift it losslessly 18dB louder so no dither can influence the noise. UMS may test if the encode suffers by its pure loudness or by other things when brought back to higher levels.
Edit: Anyone noticed the left channel only being 15bit in that sample from he beginning? No wonder lossywav can't throw away much. Very srange to have it only on one channel.
+1 with a bullet. Based on the evidence before us, let's call this snippet what it is, which is a badly made recording, and waste no more time with it. If I had some assurance that the track achieved levels within a few dB of FS at some later point, then of course it would become interesting again.
By peaking at 22 dB below FS, it sacrifices more than 3 bits worth of resolution.
Killler track and poorly made recording should be two different categories.
You didn't download it, did you? It's obviously a quiet moment in a full range classical recording. Whether it peaks at full scale I couldn't say, but it obviously goes louder than this. Maybe UMS can confirm the track and album peaks.
The only "badly made" issue is the 15bits in the left channel. That's quite weird, and probably audible. Truncating or dithering the right channel to 15 bits is audible (see above), though that would be an insane listening level for the whole piece (I assume).
Cheers,
David.
I uploaded a 3 bit louder sample for UMS to test if the problems he hears are still the same. This may clarify if it is only because of its silence or also other factors.
http://www.hydrogenaudio.org/forums/index....t=0#entry845676 (http://www.hydrogenaudio.org/forums/index.php?showtopic=102681&st=0#entry845676)
You didn't download it, did you?
Of course I did! Both files.
It's obviously a quiet moment in a full range classical recording.
Of course it is!
Whether it peaks at full scale I couldn't say, but it obviously goes louder than this. Maybe UMS can confirm the track and album peaks.
That's my point.
The only "badly made" issue is the 15bits in the left channel. That's quite weird, and probably audible. Truncating or dithering the right channel to 15 bits is audible (see above), though that would be an insane listening level for the whole piece (I assume).
15 bits in the left channel? Which sample, what time offset in the track?
Please read the entire discussion. We already had this conversation and the peak value was revealed.
FFS!
Please read the entire discussion. We already had this conversation and the peak value was revealed.
FFS!
Did that too. Seems like an obvious disqualifiying fact was being overlooked for some reason.
The clip is also pathological in that its background noise level is something like less than 30 dB below its peak level.
Please read the entire discussion. We already had this conversation and the peak value was revealed.
FFS!
Did that too. Seems like an obvious disqualifiying fact was being overlooked for some reason.
The clip is also pathological in that its background noise level is something like less than 30 dB below its peak level.
Classical recordings often present problems not found in pop music. This is a live recording in a hall, so that background noise is ambient noise floor, not recording equipment absolute noise floor. Some halls are very bad in this respect. That said, there are a LOT of pieces with very extended very soft passages. I've never yet heard the beginning of the second (slow) movement of Bartok's Second Piano Concerto done in a way that translates to a home listening system. There's an extended passage before the last section of Firebird Suite by Stravinsky that is vanishingly soft--yet perfectly audible in the concert hall if you're sitting there.
It's conceivable that classical labels could choose to normalize to -1.0db instead of -6.0, but I would bet if that range were to become standard, classical recording engineers would respond by preserving even more of the original dynamic range, which is extraordinary in comparison to pop. Your soft passages would be just as soft.
In other news, I found a slight DC offset which I believe is actually on the disk, don't know how that might interact with the treatments.
It's ironic how mp3 is capable of exceeding the dynamic range of CDDA.
We already had this conversation and the peak value was revealed.
My ability to forget things exceeds my ability to read things
It's conceivable that classical labels could choose to normalize to -1.0db instead of -6.0, but I would bet if that range were to become standard...
There is no standard. Most CDs, popular and classical, are peak normalised to nearly 0dB FS. Normalising to -6dB for a wide range classical CD is a strange choice, though some people worry about distortion in lesser equipment when getting near full scale with "pure" classical music...
http://www.audiomisc.co.uk/BBC/iPlayerRulesOK/Page2.html (http://www.audiomisc.co.uk/BBC/iPlayerRulesOK/Page2.html)
Cheers,
David.
We already had this conversation and the peak value was revealed.
My ability to forget things exceeds my ability to read things
It's conceivable that classical labels could choose to normalize to -1.0db instead of -6.0, but I would bet if that range were to become standard...
There is no standard. Most CDs, popular and classical, are peak normalised to nearly 0dB FS. Normalising to -6dB for a wide range classical CD is a strange choice, though some people worry about distortion in lesser equipment when getting near full scale with "pure" classical music...
http://www.audiomisc.co.uk/BBC/iPlayerRulesOK/Page2.html (http://www.audiomisc.co.uk/BBC/iPlayerRulesOK/Page2.html)
Cheers,
David.
I may have characterized this recording incorrectly. In the full orchestra segment here, a huge crescendo in a huge orchestra, the peak for the long clip hit -5.7 db. But it's entirely possible there's one or more points elsewhere in the full 40 minutes plus piece that approaches 0 db.
This topic has been on my mind lately and I have a few questions. But first I would like to welcome UMS to the community and high five him because his "50 plus" ears are better than my 35 year old ears, and also for tolerating far more doubt and scrutiny than I would have by repeatedly replicating his test results. I'm going to loosely quote him and greynol "trained to listen with excrutiating detail" "not enough to put myself through another abx session" and "probably not be able to abx in a car".
What I was wondering was if either of you feel you could abx short 5 or 10 second bursts.
Not asking for more tests, just an opinion.
If not, considering the entire purpose of lossy compression is portability with the best sq possible considering you are ripping out 4/5 of the audio signal, hasn't the encoder done its job? That being small files for portable or less than optimal conditions? I guess what I am saying is that offshoots or fiddling with settings are pointless because lossy is lossy and under intense scrutiny isn't the same but in less than optimal does its job just fine.
This topic has been on my mind lately and I have a few questions. But first I would like to welcome UMS to the community and high five him because his "50 plus" ears are better than my 35 year old ears, and also for tolerating far more doubt and scrutiny than I would have by repeatedly replicating his test results. I'm going to loosely quote him and greynol "trained to listen with excrutiating detail" "not enough to put myself through another abx session" and "probably not be able to abx in a car".
What I was wondering was if either of you feel you could abx short 5 or 10 second bursts.
Not asking for more tests, just an opinion.
If not, considering the entire purpose of lossy compression is portability with the best sq possible considering you are ripping out 4/5 of the audio signal, hasn't the encoder done its job? That being small files for portable or less than optimal conditions? I guess what I am saying is that offshoots or fiddling with settings are pointless because lossy is lossy and under intense scrutiny isn't the same but in less than optimal does its job just fine.
As I understand the above, I think I agree pretty much in its entirety. The MP3 was designed originally as an accompanying codec of audio stream for MPEG-1 video; as you imply, audiophile was never its intent. It has done just fine on its intended platforms, which I would identify as iPod and associated non-Apple variants: mobile devices with lower-quality earphones used in high-ambient-noise environments.
I expect the relentless march of Moore's Law (and similar "laws" of computing) to gradually remove the need for lossy multimedia compression. I know some people carry around thousands of songs, but I never have, and even that number will be possible in lossless formats relatively soon, probably on fingernail-size devices.
I originally jumped in because I had seen several blanket statements (made on other forums), that no one could distinguish 256 kbs MP3 or higher from Redbook. That kind of claim gets my attention, so my research question goes to the point of whether it's possible at all (ummm, yes), which is distinct from how useful that finding really is (not much, in the majority of cases).
I typically use 5 or 10 second segments out of a test file in Foobar anyway, so I think the answer to the question above on ABX is "yes", but maybe you mean another context.
It's ironic how mp3 is capable of exceeding the dynamic range of CDDA.
I wonder if anyone has ever found a very dynamic 24 bit classical (or other) piece where the MP3 version made from 24 bit source is transparent, but dithering down to 16 bit is not.
I expect the relentless march of Moore's Law (and similar "laws" of computing) to gradually remove the need for lossy multimedia compression. I know some people carry around thousands of songs, but I never have, and even that number will be possible in lossless formats relatively soon, probably on fingernail-size devices.
I think politics if nothing else will keep lossless from going to mainstream again. No one wants to sell lossless audio because its too easy to change the format...
I originally jumped in because I had seen several blanket statements (made on other forums), that no one could distinguish 256 kbs MP3 or higher from Redbook. That kind of claim gets my attention, so my research question goes to the point of whether it's possible at all (ummm, yes), which is distinct from how useful that finding really is (not much, in the majority of cases).
I would say that most people cannot, at present, ABX almost any common music samples for good 256k MP3. However, with practice, experience and patience, almost anyone can learn to. MP3 is quite good, but once you learn the flaws and what to listen for, you can often ABX tracks.
FWIW I prefer to not get quite so good at it that I notice these things without trying
The MP3 was designed originally as an accompanying codec of audio stream for MPEG-1 video; as you imply, audiophile was never its intent. It has done just fine on its intended platforms, which I would identify as iPod and associated non-Apple variants: mobile devices with lower-quality earphones used in high-ambient-noise environments.
Intended platforms? Maybe now, but not originally, which I am sure you know but which the proximity of your statement about its origins risks implying to readers with less knowledge of the history. PMPs were a dream on the horizon for most people when MP3 first emerged, so they could not cause its existence. Separate from whatever the Committee originally intended, MP3 originally got used predominantly on home PCs as an easier way to store song-length volumes of audio on the relatively small disks available then and/or to transfer them over the poor internet connections of that age. It precipitated the popularity of portable players (of which the iPod was definitely not the first, to which all others are merely “associated”), not the other way around.
I actually thought the initial purpose of mp3 was transporting relatively high quality audio over dual ISDN lines, hence the 128 kbps quality that used to be so common.
lower-quality earphones used in high-ambient-noise environments.
That's a bit over the top.
My turn: let's use a magnifying glass to look at the sun.
Before we get on our high horses, I'm sure any of us can find samples that you can't ABX.
Arnold got it right: this sample is pathological.
I'm sorry, but this generalization about MP3 is no less silly than the statements this discussion was set out to disprove.
I expect the relentless march of Moore's Law (and similar "laws" of computing) to gradually remove the need for lossy multimedia compression.
This phrase periodically appears in internet already for years.
The tech moves to lighter,thinner devices, hence additional restrictons for storage capacity.
We have connections of a tons of Mbits and still use jpeg for image compression. The amount of content and image/video resolution grow fast.
"Gradually remove the need for lossy multimedia compression"? Not even close.
It's ironic how mp3 is capable of exceeding the dynamic range of CDDA.
I wonder if anyone has ever found a very dynamic 24 bit classical (or other) piece where the MP3 version made from 24 bit source is transparent, but dithering down to 16 bit is not.
I may have a candidate for such a test.
I just came from a performance of Beethoven's Meeresstille und glückliche Fahrt (Calm Sea and Prosperous Voyage). Picture a chorus of 160, starting off so softly that for several seconds you have to strain to determine if they have even begun. Eventually they and the orchestra reach full volume.
I have never hear a piece with such a wide dynamic range.
Arnold got it right: this sample is pathological.
Why pathological ? Classical music is full of quiet passages so I don't think this Ravel is exceptional.
Is there any indication that MP3 transparency is reduced with low level music (assuming normal playback level) ?
Is there any indication that MP3 transparency is reduced with low level music (assuming normal playback level) ?
Before the floating ath was introduced into lame it was pretty abysmal in VBR mode (if you turned the volume up to be able to hear it). Lame 3.81...
http://mp3decoders.mp3-tech.org/lsb.html (http://mp3decoders.mp3-tech.org/lsb.html)
http://mp3decoders.mp3-tech.org/encoders.html#lame (http://mp3decoders.mp3-tech.org/encoders.html#lame)
Cheers,
David.
(assuming normal playback level)
This practical stipulation is the card that will likely make the entire house topple.
Is anyone aware of any low-level passages being used as killer samples, or aware of them being offered as evidence indicating a deficiency in the format in the last decade (until now)?
If no then why?
The MP3 was designed originally as an accompanying codec of audio stream for MPEG-1 video; as you imply, audiophile was never its intent. It has done just fine on its intended platforms, which I would identify as iPod and associated non-Apple variants: mobile devices with lower-quality earphones used in high-ambient-noise environments.
Intended platforms? Maybe now, but not originally, which I am sure you know but which the proximity of your statement about its origins risks implying to readers with less knowledge of the history. PMPs were a dream on the horizon for most people when MP3 first emerged, so they could not cause its existence. Separate from whatever the Committee originally intended, MP3 originally got used predominantly on home PCs as an easier way to store song-length volumes of audio on the relatively small disks available then and/or to transfer them over the poor internet connections of that age. It precipitated the popularity of portable players (of which the iPod was definitely not the first, to which all others are merely “associated”), not the other way around.
Yes, with "intended platforms" I was referring to common usage by consumers, not "originally intended" by developers and corporations. A paper I have published in an IS journal discusses the business development you sketch above, the appropriation of a format by consumers, and its subsequent (unwilling) adoption by the recording distribution industry.
lower-quality earphones used in high-ambient-noise environments.
That's a bit over the top.
Really? That's how I see people using their music. Maybe they are also sitting down in front of a stereo to listen to a great new album release, but that's not what they describe to me. And the earphones they use are the ones that came with the device. This isn't supposed to be a universal claim, of course. It describes the preponderance of uses that I see and which my students describe to me. And it's far less about MP3 than it is about mobile technology per se. This is just how I see people using music at present.
My turn: let's use a magnifying glass to look at the sun.
Before we get on our high horses, I'm sure any of us can find samples that you can't ABX.
These are what I'm interested in. If anyone has suggestions, I would like to try them. Just name the piece, and I'll get copies of the tracks and LAME them myself.
Arnold got it right: this sample is pathological.
I'm sorry, but this generalization about MP3 is no less silly than the statements this discussion was set out to disprove.
(assuming normal playback level)
This practical stipulation is the card that will likely make the entire house topple.
"house topple"? Not sure what the "house" is here, but I'm happy to try out any particular protocol that is offered.
Is anyone aware of any low-level passages being used as killer samples, or aware of them being offered as evidence indicating a deficiency in the format in the last decade (until now)?
If no then why?
This is a quoting style I have not seen before.
The house here is the ability to ABX these files. Under non-standard conditions (such as severely boosted volume), a lot more is ABX-able, which may lead people to believe things that don't hold at all in practical situations.
I expect the relentless march of Moore's Law (and similar "laws" of computing) to gradually remove the need for lossy multimedia compression.
This phrase periodically appears in internet already for years.
The tech moves to lighter,thinner devices, hence additional restrictons for storage capacity.
Yes, but eventually they run up against form factor restrictions instead of storage ones. Do you want an MP3 player smaller than your fingernail? We will be able to make one, but the market will not need one. And then there's the problem of driving headphones, which also does not scale down as easily. The batteries are not shrinking at a rate commensurate with Moore's Law.
We have connections of a tons of Mbits and still use jpeg for image compression. The amount of content and image/video resolution grow fast.
"Gradually remove the need for lossy multimedia compression"? Not even close.
This is a quoting style I have not seen before.
The house here is the ability to ABX these files. Under non-standard conditions (such as severely boosted volume), a lot more is ABX-able, which may lead people to believe things that don't hold at all in practical situations.
I don't know about "believe", but I made a similar point in post #68:
"For the target market and usage of MP3 generally, I'm coming to the opinion that the sort of ABX-ing I'm doing is mostly beside the point."
The belief addressed in my research question is just whether or not ABX-ing is possible at all [WITHOUT severely boosting volume, but WITH good listening equipment properly used, in a relatively good listening environment].
Yes its possible with some samples . Lossy is some compromise that is hopefully inaudible or not annoying at reasonable bitrates. I think you may want to leave lossy compression altogether - get a player (or phone) that support lossless. Its easy to do these days.