Skip to main content

Topic: ABX'ed AAC 128 VBR (log posted). Angry :( (Read 13458 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.
  • Porcus
  • [*][*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #25
As to the format discussion, which is slightly off-topic here:

- any lossless format can be replaced, at least, provided it isn't DRM'ed to the extent that it cannot be copied in the digital domain. It isn't unlikely that the fruity company might succeed with a new format to carry their miraculous hi-rez streams which will magically resolve the complex problems of 44.1/16 audio (except the ones along the real axis, which would have been addressed if I were dictator :-P )

- an MP3 is an MP3 – it may be encapsulated in a container, but it is unlikely that anyone will be able to improve over the MPEG stream until storage and bandwidth makes it uninteresting to anyone but the “because it is possible!” geeks. Which is, I guess, already.


  • C.R.Helmrich
  • [*][*][*][*][*]
  • Developer
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #26
C.R.Helmrich, did iTunes upgrade to 256kbps just because people were whining? Not sarcasm, I'd just like to understand how crazy marketing is.

I don't work for Apple, but I know they did listening tests just like Carsi did and found they needed to go to ~256 kbps stereo to be transparent even on the most difficult material they could find. Which was a great move in my opinion. It's plain obvious that 128 kbps AAC isn't transparent on several items. Neither is Opus, nor WMA, nor Vorbis, nor any future codec. The more relevant question is how close to transparency can you get at that bitrate on non-transparent items.

Quote from: jensend link=msg=0 date=
at some point between 12 and 20 years from now AAC will likely look as outdated as MP2 does today. etc.

Talking about meaningless baloney here... MP2 was designed in the mid 80s. Doesn't even use an MDCT yet. MP3 was designed around 1990 and clearly beat it. AAC was designed around the mid 90s and beat MP3, but less clearly. xHE-AAC and Opus at high bitrates are marginally better than AAC (if at all, thanks for helping out, Igor) and designed almost 15 years after AAC. Do you really think codec improvements are linear? In what way can a new codec improve? Quality? Most unlikely, marginally at most. Speed/efficiency? Most unlikely, marginally at most. For the record, I'm working on the next generation of perceptual coding 50 hours a week. So don't tell me what I write is meaningless baloney.

Chris
  • Last Edit: 29 November, 2012, 05:36:27 AM by C.R.Helmrich
If I don't reply to your reply, it means I agree with you.

  • Carsi
  • [*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #27
But today I ABX'ed Rammstein - Sonne and got 20/20. I'm so mad, I need to re-rip everything again


It sounds like you found a *particularly* bad problem sample here. Methinks a lot of HA regulars might want to listen to it (hopefully leading to improved encoders in the future). Could you upload a 15-30 second clip of the track to the Uploads forum?


Sure, I could do that, but isn't that against the forum rules? After all it's copyrighted material... If not, then I'll upload straight away, I'd love to hear what people on this forum think about this sample.

  • pdq
  • [*][*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #28
But today I ABX'ed Rammstein - Sonne and got 20/20. I'm so mad, I need to re-rip everything again


It sounds like you found a *particularly* bad problem sample here. Methinks a lot of HA regulars might want to listen to it (hopefully leading to improved encoders in the future). Could you upload a 15-30 second clip of the track to the Uploads forum?


Sure, I could do that, but isn't that against the forum rules? After all it's copyrighted material... If not, then I'll upload straight away, I'd love to hear what people on this forum think about this sample.

Thirty seconds or less is generally accepted as falling under the "fair use" category, and so is allowed under forum rules.

  • Carsi
  • [*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #29
Ok so I uploaded the sample in the uploads section:

http://www.hydrogenaudio.org/forums/index....showtopic=98111

Let me hear what you think.

  • jensend
  • [*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #30
For the record, I'm working on the next generation of perceptual coding 50 hours a week. So don't tell me what I write is meaningless baloney.
I'm quite completely aware of the historical facts, of your job, your ideological commitments, and your self-important bull. Trying to pull authority and telling people off just because "I'm with FhG, I'm soooo important" doesn't win you any points.

When you say "(xHE-)AAC and Opus are so close to the theoretical maximum in compression" that single statement is demonstrably meaningless baloney, and that's the only thing you said that I singled out as being simply baloney. No theory gives anything near objective convincing support to the claim that no future codec can achieve a markedly superior bitrate-quality curve on normal audio. Adding the qualifier "with reasonable encoding and decoding speed" doesn't help, since new algorithms, along with Moore's law, quickly bring many "unreasonable" methods and optimization problems into the realm of realistic possibility.

Again, back when MP2 was being put together, people made claims that perceptual entropy limits meant that nothing could ever do much better than MP2. Also, the kind of calculations done by a modern AAC encoder would have seemed well outside the limits of reasonable encoding speed in the late 80s.

No, I don't think codec improvements are linear. Nothing I said suggests that. Of course there's diminishing returns to effort. Nor am I saying that Opus or any other codec currently under development is going to be what obsoletes AAC. Anything which could possibly deliver that kind of an improvement will be substantially different from existing codecs and would be at best in extremely early experimental stages at this point, like Monty's chirp tracking Ghost work. (Heaven only knows whether that particular effort will lead, a decade or so down the road, to both a format which allows substantial bitrate/quality improvements and algorithms that can do the required encoder-side work with the required precision and speed, or whether it will prove to be a dead end. But it does seem to exemplify the kind of conceptual departure and exploration that may be required.)
  • Last Edit: 29 November, 2012, 11:52:27 AM by jensend

  • BFG
  • [*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #31
The banter between jensend and C.R. has me curious - I wonder what a theoretical "best possible" lossy encoding algorithm would look like, when transparency and space efficiency are the only two factors of consideration (i.e. 100% transparency for all listeners for all samples in the smallest possible package is the only goal; encode and decode times, compatibility with current hardware/software, and all other factors are irrelevant, as it is assumed that future tech would catch up).  It almost makes me want to dust off my programming skills.

And yes, I know "100% transparency for all listeners" is probably never achievable in a lossy codec.  But this is just theoretical.

  • ExUser
  • [*][*][*][*][*]
  • Read-only
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #32
I'm quite completely aware of the historical facts, of your job, your ideological commitments, and your self-important bull. Trying to pull authority and telling people off just because "I'm with FhG, I'm soooo important" doesn't win you any points.
While philosophically speaking, arguments from authority are fallacious, Mr. Helmrich is simply backing up his assertions with his area of expertise. If you would kindly grace us with a reason we should trust your opinion over that of a legitimate professional involved in reserarch, that would be much appreciated.

As a moderator, I'd also appreciate it if the tone of this discussion became less abusive.

While we don't know where the next improvements in lossy audio are going to come from, the fact is that despite many different implementations, we seem to be unable to really push the boundaries much further than where they are today. Even the people pushing the boundaries are trying to tell you that. This doesn't preclude the possibility of a huge breakthrough, but I think such a huge breakthrough is unlikely any time soon.

  • skamp
  • [*][*][*][*][*]
  • Developer
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #33
And yes, I know "100% transparency for all listeners" is probably never achievable in a lossy codec.


I know it can't compete on bitrate, but has lossyWAV been reported to be non-transparent lately? At quality setting "standard" and above, that is.
See my profile for measurements, tools and recommendations.

  • jensend
  • [*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #34
If you would kindly grace us with a reason we should trust your opinion over that of a legitimate professional involved in reserarch, that would be much appreciated.
The reason is that this "legitimate professional" has made extraordinarily sweeping universal claims without any evidence to back them up.

In my first post in this thread I made two remarks that should be entirely uncontroversial: someday there will be superior formats, and devices most likely won't play AAC forever. His arrogant dismissive "Both wrong." means that he's claiming both that there never will be any format superior to AAC and that all future devices will play AAC until the heat death of the universe. There is absolutely no good reason to believe anyone who says those kinds of things, no matter how many hours per week they're being paid to do exactly what they are saying is impossible.

We can bicker about our predictions about the magnitude of improvement future encoders will show a given number of years from now, or about how widespread backward compatibility with AAC will be a given number of years from now. I continue to think the claim that it'll be longer than 20 years before there's either a) any substantial improvements over today's AAC encoders or b) any devices which are no longer compatible with AAC is entirely incongruous with both history and theory, and as far as his "authority" goes, I'll remind you of Clarke's first law. But bickering about particular predictions is beside the point.

All that's required for my argument that using AAC for archival is non-optimal is that someday someone might want to re-encode their audio with another lossy encoder because of either an improvement in the bitrate-quality curve or because they own a new device which no longer supports AAC. The claim that these are impossibilities is unreasonable, no matter who it's coming from.

  • yourlord
  • [*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #35
I hate to interrupt the argument by going back on topic, but is that wav sample the encoded output, or the original?
Either way, can you upload a lossless copy of the encoded output or the original, whichever we don't already have..

I'm at work and don't have any tools here to really test with, but I'd like an original and a copy that suffers the problem he's describing to play with..

  • Nessuno
  • [*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #36
And yes, I know "100% transparency for all listeners" is probably never achievable in a lossy codec.  But this is just theoretical.

I think the real problem here is that as today we rely on psychoacoustic models to develop lossy codecs and on listening tests as the only way to roughly measure output errors and feedback the developers. Those models are external, although good, approximations of the real listening process, so the codec development process has an internal cause of uncertainty which prevents from asserting full transparency.

Of course medical science is not engineering as physicians don't have access to the source code...
  • Last Edit: 30 November, 2012, 04:26:25 AM by Nessuno
... I live by long distance.

  • Porcus
  • [*][*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #37
I wonder what a theoretical "best possible" lossy encoding algorithm would look like, when transparency and space efficiency are the only two factors of consideration (i.e. 100% transparency for all listeners for all samples in the smallest possible package is the only goal


The goal would likely be way more ambitious.  For a “best possible” it should be impossible to simultaneously improve both sound quality and bitrate – not only at the threshold of transparency (if that is well-defined!), but for any (reasonable) bitrate below.

But part of the argument seems to be the codec/format, and that goes beyond the single encoding algorithm. Even when an encoder can be (significantly) “improved” in the above sense, it does not necessarily mean that you have to “improve” the format in order to carry the “improved” lossy signal. Like, current LAME is better than last-century encoders – meaning, it was not necessary to introduce AAC in order to get improvements of that order of magnitude. I would consider a format to close to optimal on the above parameters, if for any “best possible” benchmark it can carry an equivalent quality without requiring much extra bitrate.



I hate to interrupt the argument by going back on topic


Oh ... THAT was beyond the pale!


  • C.R.Helmrich
  • [*][*][*][*][*]
  • Developer
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #38
His arrogant dismissive "Both wrong."

I agree, "wrong" was inappropriate. Sorry about that, jensend. Please re-interpret it as "highly unlikely".

Regarding your statement that there's no theory to support my claim "theoretical maximum in compression": there's Shannon's rate-distortion theory, where the maximum allowed distortion D is given by the theory of threshold of audibility, which in turn can be derived from the psychoacoustic theories of hearing threshold and of auditory masking (both temporally and spectrally). In my experience, encodings that clearly aren't transparent clearly violate - i.e. introduce D which is above - the threshold of audibility at some point in time, and I also think the psychoacoustic theories have become quite solid. Your opinion may differ, of course.

yourlord, sorry, yes, back to topic. To me it looks like Carsi uploaded the original. Spectrum goes to 22 kHz and looks "original".

Chris
  • Last Edit: 30 November, 2012, 07:59:12 AM by C.R.Helmrich
If I don't reply to your reply, it means I agree with you.

  • Carsi
  • [*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #39
Let's get back to topic  I uploaded the original. Anyone tested the sample yet?

  • C.R.Helmrich
  • [*][*][*][*][*]
  • Developer
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #40
Not yet, but you can save us some time by uploading the .m4a iTunes encode on which you heard the artifact(s).

Chris
If I don't reply to your reply, it means I agree with you.

  • Gainless
  • [*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #41
Let's get back to topic  I uploaded the original. Anyone tested the sample yet?


I've tested it with the recent master of Opus and the Winamp AAC encoder (both at 128 kbps), couldn't find any obvious flaws. I'm not too good at ABXing in general, though...
  • Last Edit: 30 November, 2012, 08:21:31 AM by Gainless

  • yourlord
  • [*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #42
I tested it with Vorbis down to 96kbps yesterday on some ear buds and I couldn't tell them apart using a poor man's blind test..
I'll encode the original as aac this weekend and see if I can ABX on a decent set of headphones.

  • Nessuno
  • [*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #43
@OP: exactly which parameter you passed to AAC encoder? VBR quality target? CVBR bitrate?

Anyway, I encoded at 128 CVBR (130kpbs average) and this was my first, promising attempt to ABX:

Code: [Select]
ABX Test Completed: 2012-11-30 17:54:49 +0100

Number of tests performed: 10
Number of correct answers: 8
Percentage correct:  80%

File 1 = /Users/nessuno/extempora/Sonne_chorus.m4a
File 2 = /Users/nessuno/extempora/Sonne_chorus.wav
File placement was static.

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

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


But then made some other trials and my percentage worsened, about 70% or 60% so I was clearly guessing even the first time.
... I live by long distance.

  • Alexxander
  • [*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #44
I haven't installed iTunes, but I use qaac. Encoded with foobar2000:

qaac 1.46, CoreAudioToolbox 7.9.7.9, AAC-LC Encoder, CVBR 128kbps, Quality 96

and real bitrate was 130 kbps.

This one was rather easy to ABX, at second time listening to the complete sample I spotted a problem with high-hat/cymbals (or something like that). Besides distortion with these instrument sounds I noticed a different stereo balance, but that could be caused by different level of distortion on left and right channel.

Code: [Select]
foo_abx 1.3.4 report
foobar2000 v1.1.18
2012/11/30 18:24:32

File A: C:\Users\Alexxander\Desktop\Sonne_chorus.wav
File B: C:\Users\Alexxander\Desktop\Sonne_chorus.m4a

18:24:32 : Test started.
18:25:37 : 01/01  50.0%
18:25:53 : 02/02  25.0%
18:26:04 : 03/03  12.5%
18:26:09 : 04/04  6.3%
18:26:24 : 05/05  3.1%
18:26:59 : 06/06  1.6%
18:27:16 : 07/07  0.8%
18:27:30 : 08/08  0.4%
18:27:51 : 09/09  0.2%
18:28:04 : 10/10  0.1%
18:28:08 : Test finished.

----------
Total: 10/10 (0.1%)

  • smok3
  • [*][*][*][*][*]
  • Moderator
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #45
i can not abx sample converted with

afconvert -v -f "m4af" -s 3 in.aif out.m4a
(thats default quality, fully VBRish, 126 kbps average)

(possibly due to high intolerance to specific song/band.)
PANIC: CPU 1: Cache Error (unrecoverable - dcache data) Eframe = 0x90000000208cf3b8
NOTICE - cpu 0 didn't dump TLB, may be hung

  • eahm
  • [*][*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #46
The iTunes 128kbps (High Quality) qaac setting is -a128 -q1 (ABR 128kbps Quality 1) as you can read here: https://github.com/nu774/qaac/wiki/Encoder-configuration

qaac's Quality 2 (qtaacenc's --highest) is enabled by default if variable left empty.
  • Last Edit: 30 November, 2012, 01:59:06 PM by eahm

  • smok3
  • [*][*][*][*][*]
  • Moderator
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #47
a. obviously i dont need qaac on mac, i dont use itunes, so irrelevant for me to care what defaults there
b. there was a slightly hidden suggestion to abx against true vbr (if anyone cares)
  • Last Edit: 30 November, 2012, 02:01:07 PM by smok3
PANIC: CPU 1: Cache Error (unrecoverable - dcache data) Eframe = 0x90000000208cf3b8
NOTICE - cpu 0 didn't dump TLB, may be hung

  • eahm
  • [*][*][*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #48
smok3, I was talking to the people who test, didn't really care about your post.
  • Last Edit: 30 November, 2012, 02:08:15 PM by eahm

  • jensend
  • [*][*][*]
ABX'ed AAC 128 VBR (log posted). Angry :(
Reply #49
Listening to that sample, I find it interesting that without the hint about the cymbals I probably wouldn't have noticed the difference even at considerably lower bitrates.
I agree, "wrong" was inappropriate. Sorry about that, jensend. Please re-interpret it as "highly unlikely".

Regarding your statement that there's no theory to support my claim "theoretical maximum in compression": there's Shannon's rate-distortion theory... I also think the psychoacoustic theories have become quite solid. Your opinion may differ, of course.
Well, if you'd said "highly unlikely to happen in the next 20 years" I might disagree but I wouldn't have felt the need to argue the point. And I'm sorry that I got overly worked up about this.

I agree that distortion models now seem to be reasonably accurate. And that does give you a lower bound on what compression you can get for your source model. But to say that this is the absolute lower bound for real-world music is to make an extraordinarily strong and entirely unjustified claim about your source model. In other words, it's baloney.

Are you really willing to claim that your source model contains all the prior information available when we know that a signal is, say, the PCM from someone's CD collection rather than a stream of entirely random bits? Are you also claiming that you have objectively convincing evidence of this? You might as well be saying "the asymptotic optimality of LZ for Markov sources means that the GZIP'd size of the Library of Congress is so close to its Kolmogorov complexity that no compression algorithm will do significantly better." In both cases our source models are nice and useful but certainly wrong.