IPB

Welcome Guest ( Log In | Register )

 
Reply to this topicStart new topic
WMA 192 CBR Approx. Equivalent?, Wondering how WMA 192 CBR files rate as far as sound quality?
consultant
post Nov 27 2006, 19:51
Post #1





Group: Members
Posts: 42
Joined: 24-February 04
Member No.: 12240



I'm a pseudo-audiophile. After much deliberation 3 years ago I decided to rip all my CD's using LAME with the Preset:Extreme encoding. While this setting was actually higher than I needed for the sound quality that satifies me (I'm picky but not THAT picky) I figured since I was tossing my CD's it was better to go with a very high quality encoding so there would be minimal loss in sound quality should I have to transcode later. I realize FLAC lossless would have been a better choice from that perspective but I was looking for a good compromise between sounds quality, filesize, and device/player compatiblity without having to constantly transcode. I think choosing LAME MP3 Extreme settings was a good choice.

So now I'm jsut starting to get a lot of my music from these subscription services. It came down to Rhapsody or Yahoo and I chose Yahoo because it uses 192kbit WMA files and Rhapsody uses 160kbit. I'm unfamiliar with how WMA differs from MP3 but I believe it is suppose to be a higher quality codec at the same bitrate. Some places I've seen it say WMA it about twice as good (256 WMA = 128 MP3). However it seems there are different versions of WMA?

Anyway, all I know is that these 160kbit and 192kbit WMA files on these services are CBR, not VBR. All my MP3's encoded with LAME Preset Extreme vary between 180 and 240 kbit depending on the track and they are of course VBR.

To make a short story long (which is what I did/am doing), all I'm wondering is where in the scale of quality does WMA 192 CBR fall? Is it better than the AAC 128kbit that iTunes puts out or about the same. Is it about the same as LAME MP3 Preset:Medium?

I assume it's not high enough quality to transcode to some other format to get the same sound quality with a more compact filesize (even though drive space is not an issue for me)

I guess this points out another larger issue which may have been discussed here (if so point me to the thread.)

No one really knows what the dominant audio encoding format will be 5 years from now. So the rule of thumb is to get the highest quality you can from the get go. I probably should have ripped my CD's to FLAC but LAME MP3 Preset:Extreme was still a good choice I think (it's a hell of a lot higher quality than all the MP3's that 95% of the people I know have).

So the problem is all these music download services are using lossy formats (AAC, MP3, WMA) and most of them aren't even offering the songs at high audiophile quality bitrates. I think the best service I've seen is BuyMusic.com has SOME 256kbit WMA files.

So it seems it is almost better to still just buy the CD and encode using LAME Preset:Extreme or FLAC. But these download services are SO convenient. I'm trying to assess whether I'll regret adding on to my music library using WMA 192 CBR files from here on out?

I suppose if you own a license for a track, if the service offers a higher quality version later, you'll be able to 'upgrade' But that seems sort of pathetic as you can save time if they just offered a lossless or very high quality lossy from the get go! I guess it is all about the DRM. But why not let the user choose the quality. Offer both 192kbit and 256kbit WMA of the same track. All it costs them is a few pennies for extra storage space.
Go to the top of the page
+Quote Post
pdq
post Nov 27 2006, 21:13
Post #2





Group: Members
Posts: 3305
Joined: 1-September 05
From: SE Pennsylvania
Member No.: 24233



Based on the listening tests that I have seen here over the years I have come to the following conclusions:

1) At equal bit rates there are often quality differences between different codecs, though usually the differences are small to modest.

2) At unequal bit rates the file with the higher bit rate is almost always the one with the higher quality.

Also, the only guarantee against obsolescence of your chosen codec, is to go lossless.
Go to the top of the page
+Quote Post
consultant
post Nov 27 2006, 22:34
Post #3





Group: Members
Posts: 42
Joined: 24-February 04
Member No.: 12240



Well based on some futher reading and after using ABC/HR to compare different transcodes of WMA 192kbit files, I'd say WMA 192 is very high quality, I'm guessing on par with LAME Preset:Standard. It appears the findings at www.soundexpert.info confirms this.
Go to the top of the page
+Quote Post
SebastianG
post Nov 27 2006, 23:00
Post #4





Group: Developer
Posts: 1317
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



well, soundexpert.info does use a questionable testing technique based on strong assumptions that are very likely to be invalid. You should rather do some testing on your own (normal ABX)
Go to the top of the page
+Quote Post
greynol
post Nov 27 2006, 23:15
Post #5





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



QUOTE (SebastianG @ Nov 27 2006, 14:00) *
well, soundexpert.info does use a questionable testing technique based on strong assumptions that are very likely to be invalid.

But isn't it also true that these "assumptions" will never result in a sample that sounds better than it would if the "assumptions" weren't in place?


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
eofor
post Nov 28 2006, 10:01
Post #6





Group: Members
Posts: 187
Joined: 24-March 06
Member No.: 28803



QUOTE
I'm trying to assess whether I'll regret adding on to my music library using WMA 192 CBR files from here on out


I've come to the conclusion that with lossy compression, (relative) ignorance really is bliss. By all means, do an ABX test to find out if you can spot blatantly bad quality, but the last thing you want is spend months training yourself to hear the most minute artefacts (that will be in there somewhere, even at 320 kbps), which will then ruin your listening experience forever.

Similarly, if you train yourself to spot JPG compression artefacts in digital photos, the joy of browsing through pictures on Flickr (or whatever) will never be the same.

This post has been edited by eofor: Nov 28 2006, 10:02
Go to the top of the page
+Quote Post
john33
post Nov 28 2006, 10:48
Post #7


xcLame and OggDropXPd Developer


Group: Developer
Posts: 3728
Joined: 30-September 01
From: Bracknell, UK
Member No.: 111



QUOTE (eofor @ Nov 28 2006, 10:01) *
QUOTE

I'm trying to assess whether I'll regret adding on to my music library using WMA 192 CBR files from here on out


I've come to the conclusion that with lossy compression, (relative) ignorance really is bliss. By all means, do an ABX test to find out if you can spot blatantly bad quality, but the last thing you want is spend months training yourself to hear the most minute artefacts (that will be in there somewhere, even at 320 kbps), which will then ruin your listening experience forever.

Similarly, if you train yourself to spot JPG compression artefacts in digital photos, the joy of browsing through pictures on Flickr (or whatever) will never be the same.

Precisely why you've never seen me participate in any listening tests. smile.gif Although I must confess that my aged ears probably wouldn't be much help anyway!! biggrin.gif


--------------------
John
----------------------------------------------------------------
My compiles and utilities are at http://www.rarewares.org/
Go to the top of the page
+Quote Post
Kef
post Nov 28 2006, 11:23
Post #8





Group: Members
Posts: 116
Joined: 2-December 05
From: Netherlands
Member No.: 26157



QUOTE (greynol @ Nov 28 2006, 00:15) *
QUOTE (SebastianG @ Nov 27 2006, 14:00) *
well, soundexpert.info does use a questionable testing technique based on strong assumptions that are very likely to be invalid.

But isn't it also true that these "assumptions" will never result in a sample that sounds better than it would if the "assumptions" weren't in place?


After reading the PDF about their testing methodology (http://www.soundexpert.info/articles/IGIS.pdf) I fail to see the purpose of these tests. If I understood thing correctly, SoundExpert relies on techniques as "Sound artifacts amplification" in other words, comparing input and output signals and amplifying artifacts that are not perceived under normal conditions. What is the point of such tests?

If I'm mistaken, please feel free to correct me...

/Kef

This post has been edited by Kef: Nov 28 2006, 11:27
Go to the top of the page
+Quote Post
eofor
post Nov 28 2006, 14:06
Post #9





Group: Members
Posts: 187
Joined: 24-March 06
Member No.: 28803



QUOTE (Kef @ Nov 28 2006, 11:23) *
After reading the PDF about their testing methodology (http://www.soundexpert.info/articles/IGIS.pdf) I fail to see the purpose of these tests. If I understood thing correctly, SoundExpert relies on techniques as "Sound artifacts amplification" in other words, comparing input and output signals and amplifying artifacts that are not perceived under normal conditions. What is the point of such tests?


I agree that it is fairly pointless, however I can see why it's useful if you want to compare codecs that are already 100% transparent.

QUOTE (john33 @ Nov 28 2006, 10:48) *
Precisely why you've never seen me participate in any listening tests. smile.gif Although I must confess that my aged ears probably wouldn't be much help anyway!! biggrin.gif


The 48 kbps test is really intesting though - I am absolutely amazed how good some of these codecs perform.
Go to the top of the page
+Quote Post
consultant
post Nov 28 2006, 16:37
Post #10





Group: Members
Posts: 42
Joined: 24-February 04
Member No.: 12240



Regarding the transcoding of WMA 192...

When I was doing some research I found numerous messages on various forums about people emphatically recommending NOT to transcode since it will degrade the quality, introducing more artificats. Then you see people post objective tests on this forum... by the way, what software are they using to generate those numbers (0.6, 0.4) etc?? PEAQ? I can't find any freeware version of PEAQ to download.

Anyway, what I've determined is this contention that transcoding reduced the quality is WAY overblown (mostly by audio enthusiasts who are very critical as opposed to practical.) The missing piece of information usually is what is the quality of the source file? Through ABC/HR I compared several transcodes of a WMA 192 file to MP3 (128 CBR/VBR up to 192 CBR/VBR). I couldn't tell any differences between the transcode and the original and I listended really hard.

So my conclusion is if you have high-quality source material (maybe the minimum might be AAC 128, MP3 160 VBR / 192 CBR, etc. etc.) then I think any loss of sound quality when you transcode to an equal or better encoding method is transparent for 99% of us.

This post has been edited by consultant: Nov 28 2006, 16:40
Go to the top of the page
+Quote Post
greynol
post Nov 28 2006, 20:10
Post #11





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



QUOTE (Kef @ Nov 28 2006, 02:23) *
After reading the PDF about their testing methodology (http://www.soundexpert.info/articles/IGIS.pdf) I fail to see the purpose of these tests. If I understood thing correctly, SoundExpert relies on techniques as "Sound artifacts amplification" in other words, comparing input and output signals and amplifying artifacts that are not perceived under normal conditions. What is the point of such tests?

If I'm mistaken, please feel free to correct me...
I'm sure you understand the concept and consequences of this testing better than I do.

I think the idea of the testing is to show which formats/codecs/settings give you a greater margin of transparency so that after post-processing (such as sound effects and equalization) the result will still hopefully be transparent. I think the methodology may be overly-simplified and the conclusions drawn from it suspect. Certainly there is a group of knowledgeable people here who have dismissed the testing altogether and an even larger group that scoffs at a test that is able to rank samples that are beyond transparent.

QUOTE (consultant @ Nov 28 2006, 07:37) *
Anyway, what I've determined is this contention that transcoding reduced the quality is WAY overblown (mostly by audio enthusiasts who are very critical as opposed to practical.)
I wholeheartedly agree with you! biggrin.gif
QUOTE (consultant @ Nov 28 2006, 07:37) *
So my conclusion is if you have high-quality source material (maybe the minimum might be AAC 128, MP3 160 VBR / 192 CBR, etc. etc.) then I think any loss of sound quality when you transcode to an equal or better encoding method is transparent for 99% of us.
I was recently involved in a debate over this very topic. I think transparency of a second encode is not only based on the settings but also on how large of a margin of transparency there is on the first encode. If the first encode employed the minimum settings to achieve transparency (and hence had a small margin), the second encode could easily not be transparent with an equal encoding method. You can easily prove this to yourself through blind testing.


--------------------
Your eyes cannot hear.
Go to the top of the page
+Quote Post
Mercurio
post Nov 29 2006, 00:52
Post #12





Group: Members
Posts: 104
Joined: 5-November 03
Member No.: 9669



QUOTE (greynol @ Nov 28 2006, 11:10) *
I was recently involved in a debate over this very topic. I think transparency of a second encode is not only based on the settings but also on how large of a margin of transparency there is on the first encode. If the first encode employed the minimum settings to achieve transparency (and hence had a small margin), the second encode could easily not be transparent with an equal encoding method. You can easily prove this to yourself through blind testing.


I use to think that the quality of the second encode depends on how the algorithms of the two encoders "stack", so I always prefer to ABX.

It is still not clear to me how it is possible to define a metric above transparency. Soundexperts methodology sounds strange to me because I can't imagine a "neutral" way to amplify distortions of psychoacoustic algorithms, and why this way could be indicative of the behavior of the coded file to different types of further processing... but I'm sure these were discussed here and I can't really add anything to the discussion sweat.gif
Go to the top of the page
+Quote Post
Andavari
post Nov 29 2006, 01:00
Post #13





Group: Members
Posts: 935
Joined: 3-June 02
From: USA
Member No.: 2204



QUOTE (consultant @ Nov 27 2006, 12:51) *
I figured since I was tossing my CD's it was better to go with a very high quality encoding

Going with higher quality settings as you did with LAME --preset extreme which is now -V0 is wise. However the part of tossing out your CD's (I think that's what you meant) isn't an idea I'm too fond of, and I think you'd be better off putting them in a cardboard box and taping it up to keep out dust, and then stuff them into a closet - that way should the need arise you could re-rip them again into another format without transcoding.


--------------------
Complexity of incoherent design.
Go to the top of the page
+Quote Post
ezra2323
post Nov 29 2006, 03:46
Post #14





Group: Members
Posts: 586
Joined: 17-July 02
Member No.: 2631



www.allofmp3.com is a digital music service that offers songs in Lossless quality. Around .75 per song at Lossless. The site is still legal although the US is pushing Russia to shut it down.

Personally, on an iPod - with average headphones, I think purchased iTunes at 128 kbps sound great. Those same tracks burned to an audio CD and played on a high end stereo - not as much. I think the equipment you use to listen to music matters more than the codec. It will take some decent audio equipment and sensitive ears to bring out artifacts in WMA at 192.
Go to the top of the page
+Quote Post
halb27
post Nov 29 2006, 09:16
Post #15





Group: Members
Posts: 2414
Joined: 9-October 05
From: Dormagen, Germany
Member No.: 25015



QUOTE (Mercurio @ Nov 29 2006, 01:52) *
...It is still not clear to me how it is possible to define a metric above transparency. ...

Though I admit it is not easy to decide on the exact value of the soundexpert test to me it is a valuable addition to learning about codec's behavior.
I look at it like this:

What we're out for with lossy codecs is transparency, but transparency is a soft subject. It depends on the listener, and it depends upon the listener being trained to carefully listen.
As for this it makes sense to provide for a test that makes it easy to hear artefacts.

It's true that the artefact amplification is questionable, and more so any specifique technique for amplifying artefacts. But going more general this is true for any specific listening test due to the fact that only a few samples can be tested. The test can be done very carefully, but if a certain codec performs poor on a pretty big class of samples, but no member of this class has made it into the test, this codec can come out fine in the test, but in reality isn't that good. Moreover behavior on problem samples is important too. We may not expect perfect behavior on problem samples, but we should expect good or at least adequate quality (measured from behavior of the encoder universe for this codec).

So to me any listening test is valuable, but in a restricted sense, and this holds for the soundexpert tests as well.

This post has been edited by halb27: Nov 29 2006, 09:17


--------------------
lame3100m --bCVBR 300
Go to the top of the page
+Quote Post
Gigapod
post Nov 29 2006, 11:30
Post #16





Group: Members
Posts: 152
Joined: 2-November 06
From: Yvelines, France
Member No.: 37080



QUOTE (halb27 @ Nov 29 2006, 09:16) *
...
What we're out for with lossy codecs is transparency, but transparency is a soft subject. It depends on the listener, and it depends upon the listener being trained to carefully listen.
As for this it makes sense to provide for a test that makes it easy to hear artefacts.
...
So to me any listening test is valuable, but in a restricted sense, and this holds for the soundexpert tests as well.

I quite agree with your views.
You can also check the Transparency topic in the Hydrogenaudio Knowledgebase.

Of note is the fact that, by definition, you cannot prove that any codec is transparent at any given bitrate, you can only prove (using ABX) that a codec is not transparent, at a certain bitrate, on a limited number of audio samples. Whether these audio samples are chosen or not to highlight artifacts caused by the codec being tested, is of course the choice of the tester, a choice that has to be analyzed with a critical mind when reading the results of the test.

So any argument to prove the transparency of a codec at any bitrate is a futile one.

OTOH, any argument to prove that a codec is not transparent, even when supported by experimental data (results from ABX tests perfectly conducted - a rare thing indeed) will meet with criticism, based on the necessarily limited test conditions (choice of audio samples, number of test subjects, training of test subjects, quality and variety of equipment used for playback, etc).
Go to the top of the page
+Quote Post
Serge Smirnoff
post Nov 29 2006, 14:13
Post #17





Group: Members
Posts: 370
Joined: 14-December 01
Member No.: 641



QUOTE (consultant @ Nov 27 2006, 21:51) *
... all I'm wondering is where in the scale of quality does WMA 192 CBR fall? Is it better than the AAC 128kbit that iTunes puts out or about the same. Is it about the same as LAME MP3 Preset:Medium?

My findings (if they are meaningful for you) show that wma@192cbr is definitely better than iTunes aac@128cbr and Lame mp3@V4(Medium). It is inferior to aac-family-encoders @192 only (except for iTunes) as wma quality is more dependable on samples used for testing. Unfortunately several recent outliers have just spoiled results of Nero aac@192 at SoundExpert but the defeat is temporary.

Let me also add some oil to the flame. Here is an excerpt from future article (draft) aimed to explain to Mr. Joe Average how SoundExpert testing works. This part devoted to “transparency margin” discussed above.

CODE
… such analytically computed ratings are usually located above 5th grade on impairment scale. This could be interpreted as quality margin or quality headroom of an audio device because the artifacts are beyond the threshold of human audibility. You may ask what the purpose of the margin is if sound artifacts are inaudible already. There are at least four reasons why this is important:

•    In general case perceived audio quality of a device/technology depends on sound samples used for testing. Theoretically a listening test has to be performed with infinite number of sound samples in order to prove for sure that tested device will not produce unexpected “surprises” on real-world audio material. In practice a limited set of typical or problem (“killer”) sound samples is used. Then testing results are just generalized on all audio. Obviously quality margin makes that generalization more grounded and lessens the probability of getting artifacts on audio material not used during the test.

•    Very often audio devices/technologies are used in chains – connected one after another. In most cases this accumulates sound degradation throughout the chain. Quality margin of each device is highly desirable to lessen overall distortion level.

•    Such post processors as equalizers, spatializers, SRS and many others usually reveal sound artifacts inaudible in “normal” cases. Some quality headroom helps to use all those popular sound enhancements safely without danger of discovering drawbacks of other audio components.

•    Human hearing abilities differ from person to person. Averaged results of any listening test have to be applied with great caution to someone’s particular situation especially if that someone has “golden ears”. Such person needs audio equipment with sufficient quality margin in order to be satisfied.

Sound quality margin is not something completely new. Well known technical audio parameter – THD is used quite similarly:  measured on pure sine wave and corresponding to perceived audio quality not very well it have to be as low as possible – far beyond human abilities to hear such low distortions of pure waves.


--------------------
keeping audio clear together - soundexpert.org
Go to the top of the page
+Quote Post
Mercurio
post Nov 29 2006, 15:52
Post #18





Group: Members
Posts: 104
Joined: 5-November 03
Member No.: 9669



Serge, even if I'm going off topic, since this excerpt is for the Average Joe, maybe you may like a feedback from an Average Joe like me.

I still don't understand how can you define a metric "above transparency" for a psychoacoustic codec.

Also I really think that the THD example can give a wrong information to the reader, since it is defined without regarding listener's perception.

Last, while it is very difficult to write a codec that outperform the others in "traditional" abx tests, is it hard to write one that exploits soundexpert's tests at a given bitrate? (even simply preprocessing the audio signal and then using a standard codec)

QUOTE
Obviously quality margin makes that generalization more grounded and lessens the probability of getting artifacts on audio material not used during the test.


I'm sorry to me it is not so obvious, even if I try to assume that I understand what quality margin is. Maybe you should explain a bit better here.

This post has been edited by Mercurio: Nov 29 2006, 16:04
Go to the top of the page
+Quote Post
Serge Smirnoff
post Nov 29 2006, 16:44
Post #19





Group: Members
Posts: 370
Joined: 14-December 01
Member No.: 641



QUOTE (Mercurio @ Nov 29 2006, 17:52) *
I still don't understand how can you define a metric "above transparency" for a psychoacoustic codec.

For me and for SE testing method codecs are similar to any other audio processing devices like amplifiers for example. True, it doesn’t matter what principals lay behind processing nature – psychoacoustic or not. Phenomenon of transparency margin objectively exists and can be applied to all audio devices (excluding deliberately poor sounding). The problem is to find appropriate metric and measurement method. I proposed some variant. It has its pros and cons.

QUOTE
Also I really think that the THD example can give a wrong information to the reader, since it is defined without regarding listener's perception.

Yes, it was defined without regarding listener's perception but is widely used exactly with that regard.

QUOTE
Last, while it is very difficult to write a codec that outperform the others in "traditional" abx tests, is it hard to write one that exploits soundexpert's tests at a given bitrate? (even simply preprocessing the audio signal and then using a standard codec)

Sound artifacts amplification for hearing is quite the same as magnifying glass for vision. Can you cheat the glass?


--------------------
keeping audio clear together - soundexpert.org
Go to the top of the page
+Quote Post
consultant
post Nov 29 2006, 17:06
Post #20





Group: Members
Posts: 42
Joined: 24-February 04
Member No.: 12240



QUOTE (Mercurio @ Nov 28 2006, 15:52) *
I use to think that the quality of the second encode depends on how the algorithms of the two encoders "stack", so I always prefer to ABX.

It is still not clear to me how it is possible to define a metric above transparency. Soundexperts methodology sounds strange to me because I can't imagine a "neutral" way to amplify distortions of psychoacoustic algorithms, and why this way could be indicative of the behavior of the coded file to different types of further processing... but I'm sure these were discussed here and I can't really add anything to the discussion sweat.gif


I agree and therefore regret throwing 70% of my CD collection away. I did keep my really favorite CD's so at least I can re-rip them if I want. Without the cases they don't take up much room but I just hate being a pack rat. If I had to it all over again, I'd get the largest hard drive I could, rip everything to two versions, one FLAC version for archival and one with LAME preset-Standard. It would probably be wise to buy a second external hard drive to use as a backup for all the FLAC files (or back them up to a tape drive)
Go to the top of the page
+Quote Post
SebastianG
post Dec 7 2006, 03:07
Post #21





Group: Developer
Posts: 1317
Joined: 20-March 04
From: Göttingen (DE)
Member No.: 12875



QUOTE (Serge Smirnoff @ Nov 29 2006, 16:44) *
Sound artifacts amplification for hearing is quite the same as magnifying glass for vision.

No it isn't because you only amplify the difference.

A more appropriate analogy would be:
Suppose you take a picture with a digi cam of a still scene.
Move one object slightly ("movement artefact")
Take another picture from the same location.
These will look pretty much the same.
However, your testing methodology will amplify the per-pixel differences. The differences -- most prominent near the object's edges/silhouette -- will be highlighted. You'll see something that isn't in either of both images. Of course this can be detected by our eyes easily.

You try to artificially increase percibility of artifacts for understandable reasons. But you're doing it via amplifying the difference which is just a very very bad idea. What's a "correct amplification" in the above example? You could try to detect the movement and extrapolate a new picture from it. That's one option which is obviously only suited for one type of artefact -- the movement artefact. And now we all should realize that no proper artefact amplification exists or is at least very hard to implement. How we perceive things needs to be considered as well as what kinds of artefacts are involved.

What's that got to do with audio? Well, replace the term "movement" with "phase shift" and think of high frequency noise (>2 kHz) whose polarity has been inverted (flat 180° shift for all frequencies). The new signal exhibits the exact same signal energy in the exact same frequency/time regions and is very very likely to sound the same to you. Obviously your ear doesn't care about the difference. If you "amplify the difference" you just scale your signal. Scaling it naturally affects perceived loudness. So we do notice a change (in loudness). What does this prove? Difference amplification is meaningless.

Consider a signal fraction that represents a certain frequency/time subset to be a vector in R^n. Suppose that we only notice changes when the vector's length is altered (simplification) and that your original signal fraction is given by X. Then Vorbis' psychoacoustic model determins that a certain error E is tolerable as long as the length of the vector X+E, |X+E| is close to |X| which is what "noise normalization" tries to do. For similicity we choose |X+E|=|X|. There are many many solutions to this and all are bounded by |E|<=2|X|. Given the way we perceive things a proper amplification of the error would be to choose an error E' so that
1) |X+E'| = |X+E| = X (even when lowering the bitrate NN tries to fulfill this)
2) E' is a linear combination of X and E.
3) angle(X,X+E) < angle(X,X+E') (amplify the "error angle")
where angle determines the angle between two given vectors. What you're doing is to simply set E' to be a multiple of E which violates |X+E'|=|X| and invalidates "noise normalization". By setting E' to a multiple of E you are oversimplifying the way we perceive artefacts. If you oversimplify this you can't draw meaningful conclusions.

Audio coding techniques like "noise normalization", PNS, PS and SBR try to make it sound similar although they allow a possibly large difference in a domain you're more familar with. Another example would be parametric speech coding. These coders don't care about phases at all, they only generate a filtered buzz which sounds close to the original. I'd like to quote Christopher Montgomery, the man responsible for Ogg Vorbis: "F*ck the SNR" (from a private conversation during a time NN was in an early stage).
Ironically SBR performs quite well in your tests which suggests that your amplification technique does less amplification in the way we're sensible to on those kinds of artefacts than what would be appropriate.

In a nutshell: You assume the percibility of an artefact is doubled by doubling the per-sample difference. This is the strong assumption I meant your tests are based on. I hope you could follow my reasoning above for why difference amplification is a bad idea when it comes to those kinds of audio tests.

regards,
Sebastian

This post has been edited by SebastianG: Dec 7 2006, 03:34
Go to the top of the page
+Quote Post
Serge Smirnoff
post Dec 20 2006, 16:27
Post #22





Group: Members
Posts: 370
Joined: 14-December 01
Member No.: 641



QUOTE (SebastianG @ Dec 7 2006, 05:07) *
A more appropriate analogy would be...

Of course any analogy is poor in some sense. So I’ll try to answer using your one.

QUOTE
You try to artificially increase percibility of artifacts for understandable reasons. But you're doing it via amplifying the difference which is just a very very bad idea. What's a "correct amplification" in the above example? You could try to detect the movement and extrapolate a new picture from it. That's one option which is obviously only suited for one type of artefact -- the movement artefact. And now we all should realize that no proper artefact amplification exists or is at least very hard to implement. How we perceive things needs to be considered as well as what kinds of artefacts are involved.

Agree. Sensitivity of human visual (hearing) system towards different types of artifacts highly depends on the nature of those artifacts. Moving objects can be more perceptible for example than changing their colors. In fact SARTAMP amplifies signal differences (in some wise way, though; see next paragraph) but how these differences are perceived is determined in each case (tested device/sound sample) individually by means of psychometric curve construction. My findings show that functional dependence between signal difference amplification (for a device/sample combination) and human perception of those differences can be safely described with 2nd order curve of some (negative or positive) curvature. This psychometric curve aims to determine in each case how we perceive those particular artifacts. So “how we perceive things … as well as what kinds of artifacts are involved” is considered by this psychometric curve.

QUOTE
What's that got to do with audio? Well, replace the term "movement" with "phase shift" and think of high frequency noise (>2 kHz) whose polarity has been inverted (flat 180° shift for all frequencies). The new signal exhibits the exact same signal energy in the exact same frequency/time regions and is very very likely to sound the same to you. Obviously your ear doesn't care about the difference. If you "amplify the difference" you just scale your signal. Scaling it naturally affects perceived loudness. So we do notice a change (in loudness). What does this prove? Difference amplification is meaningless.

Sorry for citing my own paper:
As it was showed in [FEITEN B, “Measuring the Coding Margin of Perceptual Codecs with The Difference Signal” presented at the AES102th Convention … ] additional filtering of the resulting signal is necessary for eliminating of the frequency components, absent in output signal and thus appeared in resulting signal after subtraction. Such filtering was performed by FIR filter with the magnitude response vector, calculated according to slightly modified algorithm from [paper above].

Such filtering will also eliminate that gained “inverted noise” made by plain subtraction of the signals. In combination with very characteristic (for this case) psychometric curve the final computed quality score will be almost infinity (if the only problem with that signal was inverted noise …).

Also because of these two SARTAMP features (filtering and psy-curve construction) there are no problems with SNR and “noise normalization”. If (in your terms) error amplification is made improperly by SARTAMP (invalidating noise normalization) then unmasking of E will be more rapid (with gradually increasing difference level) than in case of “proper amplification”. As a result the psy-curve will be steeper which in turn will result in higher final score compensating this way “incorrectness” of amplification. This is in brief. But in any case the overall correctness of such hearing-wise amplification method can be verified only in standard listening tests which I hope to organize sooner or later.

QUOTE
Ironically SBR performs quite well in your tests which suggests that your amplification technique does less amplification in the way we're sensible to on those kinds of artefacts than what would be appropriate.

SoundExpert tests show that SBR @128 and @192 performs at least as good as other technologies and effectiveness of double-rate SBR increases substantially @320. Do you think it is unusual?

QUOTE
In a nutshell: You assume the percibility of an artefact is doubled by doubling the per-sample difference. This is the strong assumption I meant your tests are based on. I hope you could follow my reasoning above for why difference amplification is a bad idea when it comes to those kinds of audio tests.

The assumption you mentioned was used in early works devoted to so-called “coding margin”. Definitely it was an oversimplification of the problem of artifacts amplification as the approach didn’t consider the way human hearing perceives different types of signal artifacts. SE quality margin estimation is smarter in this sense. It uses another assumption (in your terms as well) – perceptibility of an artifact depends not only on amount of per-sample difference (measured by Difference level) but also on audibility of that particular difference (considered by psychometric curve).

Hope my reasoning was understandable as well.


--------------------
keeping audio clear together - soundexpert.org
Go to the top of the page
+Quote Post
Ivan Dimkovic
post Dec 20 2006, 19:17
Post #23


Nero MPEG4 developer


Group: Developer
Posts: 1466
Joined: 22-September 01
Member No.: 8



QUOTE
SoundExpert tests show that SBR @128 and @192 performs at least as good as other technologies and effectiveness of double-rate SBR increases substantially @320. Do you think it is unusual?


I do wink.gif

I think it is completely in disagreement with how LC-AAC and SBR work, but oh well...
Go to the top of the page
+Quote Post
pepoluan
post Dec 20 2006, 20:12
Post #24





Group: Members
Posts: 1455
Joined: 22-November 05
From: Jakarta
Member No.: 25929



In response to the OP's situation (and slightly offtopic):

I think someone ought to do a test on transcoding, i.e. which lossy encoding will transcode best to a certain target.

E.g., let's say, the target is MP3 CBR 128 kbps.

Now we start by ripping some tracks, encode e.g. to MP3 CBR 320, MP3 V0, Vorbis -q10, Vorbis -q9, WavPack Lossy 384, WavPack Lossy 320, etc. Next we transcode the high-bitrate-encoded tracks to the target. Then we compare them.


--------------------
Nobody is Perfect.
I am Nobody.

http://pandu.poluan.info
Go to the top of the page
+Quote Post

Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 21st April 2014 - 16:55