Talk:Transparency (data compression)

Method or Example
Needs a example after the first paragraph. Something like this... (please improve on the phrasing or give a better example)

"For example, when encoding an image, a group of pixels with a small colour difference may be assigned just one colour between them. This would improve the compressibility of the data. The smaller the colour difference, the more transparent the compression." —Preceding unsigned comment added by 82.7.196.141 (talk) 08:53, 10 July 2010 (UTC)

Testing
I agree about testing being difficult. Personally I am aware that my hearing is not always in the same condition - sometimes I have wax in my ears, and sometimes even signs of tinnitus.

Superficially for MP3 at 192 kbps I can't tell the difference between most CDs and the encoded version. However if I listen to the music while encoding, I can tell that there are some critical parts where there is a perceptual loss even with 192 kbps encoding. These are infrequent perhaps only a few seconds in a long piece of music, but often these are the parts where I feel "wow" in a good recording. The MP3 encoding will often remove the wow factor.

In order to do a fair test the recordings used should be good. In my opinion, many people, even some with quite impaired hearing, can tell the difference between a good recording and a poor one. Recently in a blind test I was asked about the quality of 160, 128 and 64 kbps (I think) MP3 files. I was not given access to the original recording. There was little difference between the 160 and 128 kbps encoded material. There was a harshness in the string sound (orchestral, Tchaikovsky - Serenade for Strings) which very probably did not reflect the sound of the orchestra. The low bit rate recording did not have too many obvious MP3 artefacts, and it was clear that it had a smoother, but also possibly duller frequency response. I chose that as my preferred version - to the amusement of the testers. However if I'd been given a reference recording, or been asked to do an ABX comparison I would probably had been able to do that. The point I'm making here is that the filtering of high frequency sounds in the particular source actually made it pleasanter to listen to. Also I thought that the stereo separation was better in the low bit rate version. I would have expected a really high quality sound source to have sounded smooth, but also significantly less dull than the low bit rate recording.

Where this kind of testing is carried out, the testers really do need to know what they are doing. Significant care needs to be taken when generating the encoded material, and possibly some familiarity with the music used for testing, and the nature of artefacts in the encoding process will help the testers to develop material for a valid test.

In the limit, for higher quality encoding, the differences may be very subtle and small, and doing the evaluation of subjects could be very hard. It is possible that some differences in the sounds perceived may be slight modifications and have little to do with the encoding. These could include level differences and frequency response differences. For me after reading about the MAD MP3 decoder I've been using Winamp 2.95 with the MAD decoder in preference to newer versions of Winamp (see http://en.wikipedia.org/wiki/Talk:Transparency_%28data_compression%29 ). There are articles which tell me that this should sound better than the standard decoder. I am thus pre inclined to prefer this, even if I really cannot tell the difference. I feel however that I can actually detect a difference and recordings do sound pleasanter, more real, and more involving, but this could be due to frequency changes rather than the finer resolution of the decoding due to the dithering used. I have no way of knowing this, and only an independent test would establish this in my perception. I suppose I could evaluate the frequency content of samples using different versions of Winamp to determine whether frequency modifications are a possible explanation of my preference for the MAD decoded data.

David Martland 11:01, 3 May 2007 (UTC)

There's a lot of audiophile "I can hear more than you plebs" style boasting surrounding this topic. The only valid tests are double-blind with at least a 90% confidence threshold (i.e. there's less than a 10% chance that you got it right by guessing). Anything less is just too likely to be imagination (or baseless boasting). The 90% figure is arbitrary. Obviously you can choose a higher one if you want to be more certain but anything less will probably be laughed at. 155.198.65.29 (talk) 18:13, 12 May 2009 (UTC)

Double blind testing and proof
This wiki page correctly states that: 

 "There is no way to prove whether a certain lossy compression methodology is transparent. To scientifically prove that a compression method is not transparent, double-blind tests may be useful. The ABX method is normally used."  

User Binksternet has sought to change this pre-existing definition to read as follows:

 If two different versions of a sound are given the same score in a double blind test, the two sounds are proven "transparent"—to have no subjective difference. 



The short answer to this is that it is false, and therefore violates WP:EL. The longer answer is that it belies a correct understanding of proof and exactness, and the role of what hypothesis testing does and can do. The original wiki defn appears correct. The difference is perhaps subtle, but it is important. (talk) 12:58, 5 October 2009 (UTC))


 * The reference can't possibly violate any of the guidelines about external links! It's a perfectly good reference. Pohlmann is an expert on the subject. Don't remove it. Binksternet (talk) 18:12, 5 October 2009 (UTC)


 * Also, if you wish to defend the earlier text, you will need to have a reference to back it up. The fact that it was pre-existing here doesn't count for anything in the face of a complete rebuttal in the form of an expert source. The expert: "Ken C. Pohlmann is a professor at the University of Miami in Coral Gables, Florida, and the director of the Music Engineering program..." If you find a source that states otherwise, we can come to a consensus about how to bring both viewpoints to the article. Binksternet (talk) 18:30, 5 October 2009 (UTC)


 * The current text seems very illogical, despite coming from an "expert source." --Kjoonlee 22:37, 5 October 2009 (UTC)
 * What if two samples both got 0 out of 5 points? Would they both be equal? --Kjoonlee 22:48, 5 October 2009 (UTC)

The ABX test is similar to what statisticians refer to as a duo-trio test (if I remember correctly). As it is, an ABX test is based on alternative hypotheses and null hypotheses so you can never prove your alternative hypothesis (you can't prove that there is no difference); you can only fail to prove the alternative hypothesis (you can only fail to prove that there is a difference). When you say that someone passes an ABX, that means that there is a high enough chance that it couldn't have been blind luck. This comes from statistics more than anything else. --Kjoonlee 22:47, 5 October 2009 (UTC)


 * Have you read the Pohlmann piece? Perhaps you would prefer to reword the bit about transparency based on his text. What is not okay is for unreferenced text to replace referenced. Binksternet (talk) 23:03, 5 October 2009 (UTC)


 * Here's a quick view of pages 407 and 408 of Pohlmann's book, the section about critical listening where he discusses transparency, double blind testing, ABX tests, subjective tests, and the subjective difference grade (SDG) which can result in the conclusion that transparency has been reached if the scores are the same between the hidden reference and the codec under test.
 * http://books.google.com/books?id=VZw6z9a03ikC (go to page 407.) Binksternet (talk) 23:14, 5 October 2009 (UTC)
 * Maybe I was rash to say that the ref was useless, without reading it first. But the wording (as quoted on this talk page) was very problematic, and cannot be used. Let's ignore sources for the moment (as per WP:IAR) and and focus on what's correct. --Kjoonlee 23:27, 5 October 2009 (UTC)


 * I would like to discuss rewording the bit about transparency, based on Pohlmann, and based on any other expert you care to bring forward. I'm not willing to let the article baldly state that transparency can't be determined as Pohlmann contradicts this directly. Once you have another one or more expert sources saying that it can't be achieved, we can discuss instead how the section can be written to incorporate contradictory expert views. What I don't agree with is your conclusion that this reference should be ignored. Instead, let's reword the paragraph. Binksternet (talk) 23:56, 5 October 2009 (UTC)


 * Some more references which touch upon transparency and testing:
 * Multimedia technologies: concepts, methodologies, tools, and applications, Volume 3, page 758. Mahbubur Rahman Syed, Syad Mahbubur Rahman. Idea Group Inc (IGI), 2008. ISBN 1599049538
 * Digital audio watermarking techniques and technologies: applications and benchmarks, page 268. Nedeljko Cvejic, Tapio Seppänen. Idea Group Inc (IGI), 2007. ISBN 1599045133
 * Audio signal processing and coding, age 389. Andreas Spanias, Ted Painter, Venkatraman Atti. Wiley-Interscience, 2007. ISBN 0471791474
 * Introduction to digital audio coding and standards, page 241. Marina Bosi, Richard E. Goldberg. Springer, 2003. ISBN 1402073577
 * Every one of these textbooks says that transparency can be determined between two audio codecs. What's left here is to decide how to word our paragraph about it, how to discuss the testing procedures which allow a transparency conclusion. Binksternet (talk) 01:27, 6 October 2009 (UTC)
 * Page 408 is unavailable at Google Books, but judging from what you write, I say it's either your misunderstanding or the so-called experts' misunderstanding. Any expert on perceptual coding should be aware of statistics as well, and any of them will be able to tell you that you can never prove an alternative hypothesis; you can only reject a null hypothesis with a predefined certainty. --Kjoonlee 23:03, 6 October 2009 (UTC)
 * Correction: You can only reject or fail to reject a null hypotheis. --Kjoonlee 23:26, 6 October 2009 (UTC)


 * That's great... if "any expert" is aware, then it should be very easy for you to find a cite. Binksternet (talk) 01:18, 7 October 2009 (UTC)

 

Kjoon is quite correct. You can only reject or fail to reject a null hypothesis. This is standard first year statistics. 

In hypothesis testing, the null hypothesis is assumed to be correct, unless it is proven otherwise. Most standard 1st year texts make the analogy with someone being on trial for a crime. The null hypothesis is that the person on trial is assumed to be innocent, unless they are proven guilty. By analogy, the two sounds are assumed to be the same (to be transparent), unless they are proven to be different. A court case never proves that someone is innocent ... it either finds that someone is guilty, or it fails to find them guilty. It does not prove them innocent. See, for example, Judge, Hill, Griffiths, Lutkepohl and Lee, Introduction to the Theory and Practice of Econometrics (1988, pp.93-95).

Or, see Spanos (1999), Probability theory and statistical inference (pp.699), who tries to put it into lay terms as follows:

"The above solution amounts to a convention which views the Type 1 error as much more serious and thus the null and alternative hypotheses are treated asymmetrically. By fixing the Type 1 error to be a small number, say 0.01, we view it as much more important than the Type 2 error. Hence, we consider the mistake to reject hte null when valid to be much more serious than accepting the null when false. An emotionally charged way to rationalize this conventional choice is in terms of the analogy with a criminal offense trial. The jury in a criminal offense trial are instructed by the judge to find the defendant not guilty unless they have been convinced 'beyond any reasonable doubt' by the evidence presented in court during the deliberations. That is, they are instructed to choose between:

H0: not guilty and   H1:  guilty

The clause beyond any reasonable doubt amounts to fixing the Type 1 error to a very small value. This is designed to protect innocent people from being found guilty. This strategy, however, increases the risk of 'letting a lot of crooks off the hook' .... " (Spanos 1999) --

In exactly the same way, tests for transparency assume that the two sounds are the same, and then look for evidence to prove otherwise. If such evidence is absolutely conclusive, we have proof that they are different. But if the evidence that they are different is only so-so, we fail to reject the null hypothesis (that they are the same), and that is all. Nothing is proven.

I am sure your intentions are honourable. However, it is really most inappropriate for anyone to go and reverse the content of an encyclopedia to a logically inconsistent position, when they are unfamiliar with the technical nature of the material, and appear to be formally unqualified to write about the technical subject matter, and then expect other people to have to explain the literature to you.

I would also add, that in academia, it is usually considered bad form to list references on a topic, unless those references have actually been used to create the article. (KrodMandooon (talk) 11:40, 7 October 2009 (UTC))


 * My intention is to use those references in the article. What references were used in this article? None are apparent. Binksternet (talk) 13:21, 7 October 2009 (UTC)


 * It appears to me that there is a difference in the industry between theory and practice. In theory, transparency is not possible, yet in practice, transparency is noted if certain conditions apply. Binksternet (talk) 14:28, 8 October 2009 (UTC)
 * Transparency itself is very possible; it's just that it's impossible to prove whether some lossy compression process will be transparent for all people with different ears for all kinds of wide-ranging input at all compression quality levels etc etc. Yet each person can find personal examples where compression artifacts can't be noticed, where "personal" transparency has been achieved.
 * Let's say a person listened to three samples, A, B, and C. A is the original and B and C are compressed samples. If a person rates both B and C as 5.0 out of 5.0, then B and C may well be transparent for that person; we just can't prove it, that's all. --Kjoonlee 20:40, 8 October 2009 (UTC)


 * And yet, in the expert literature, there are examples where sound samples are said to be transparent with respect to each other, where the researchers have tested enough people enough times to call their codec transparent. The term appears to be used much more casually than its strict definition. This casual usage, since it is present in the lit, can be present in the article. Binksternet (talk) 02:46, 9 October 2009 (UTC)


 * Quite the opposite ...

Binksternet's admission that his references are erroneous is accepted. It would now be appropriate to remove those references from wiki before they cause further contamination. Thanks. (KrodMandooon (talk) 12:43, 9 October 2009 (UTC))


 * I missed my "admission" that my "references are erroneous..." Where was that? I am instead seeing that you are deleting valid references as well as deleting maintenance tags such as fact tags (citation needed tags) without fixing the indicated problems. Deleting rightful maintenance tags is vandalism. Deleting good references and replacing them with no references is vandalism. I don't think you have a leg to stand on until you a) find references to back up the challenged facts and b) find references which contradict the ones in the article, at which point we work together to decide how the contradictory expert viewpoints can be represented fairly. Binksternet (talk) 20:19, 9 October 2009 (UTC)
 * I have yet to see another definition of transparency. If the researchers are using the term sloppily, then it just means they're sloppy scholars. If they say transparency can be proved, then they should be beaten over the head with a Statistics 101 textbook. --Kjoonlee 23:39, 9 October 2009 (UTC)
 * If you want to add tags then be my guest. But don't add refs that don't contribute any new stuff to the article, please. We don't think they're good references if they contain errors about transparency. --Kjoonlee 23:40, 9 October 2009 (UTC)

It doesn't matter whether you think they are good references or not. The fact remains that he has a reference and you do not. Per WP:V, we retain only verifiable information. "The threshold for inclusion is verifiability, not truth." If what you say is true, you can find sources to back it up. "The burden of evidence lies with the editor who adds or restores material."

Your use of WP:IAR is invalid, as it violates not only the text but the spirit of verifiability. WP:V is specifically designed for cases like this. Even your reason for invoking WP:IAR is against policy. And your actions very likely scared off a good editor who did follow policy, even if he lacked the ability to defend himself using it. And your last statement violated WP:OWN

I will do a little bit of research to check your assertions, and, if I find them, I will alter the text accordingly to use both statements, which is what should have happened in the first place. — trlkly 19:05, 21 October 2011 (UTC)


 * Forgot to mention: you also violated WP:OR, as did your colleague. And both of you violated WP:OWN much more often than I thought, claiming that only experts such as yourselves could edit the article. That is so against the spirit of Wikipedia that I am having a hard time believing that you are familiar at all with how things work. Your citation of WP:EL, which only applies to extra links added at the end of the article, doesn't help there, either.


 * I hope that, in the 2 years since you read this, you have familiarized yourself with how Wikipedia works. All editors are equal, and expertise is irrelevant except in deciding which source to believe. — trlkly 19:19, 21 October 2011 (UTC)
 * I just find it laughable that someone with no knowledge can ignore verifiable facts if they have a source they can twist to their liking. --Kjoonlee 14:24, 2 April 2013 (UTC)
 * I also find it laughable that then that would lead to accusations of violations of policies/guidelines - for trying to remove verifiably false information, more than anything... --Kjoonlee 14:28, 2 April 2013 (UTC)

Acuity
Age of listener, and other factors (noisy work environment, frequent firearms usage or loud music environments w/o ear protection, etc.) are a major factor, too, but not mentioned. With treble especially, younger, less pounded ears are far, far more sensitive (enough so that there are are joke Android/iOS apps called "teen annoyers" that make startling high-pitched sounds people older than about late 20s can't hear at all. This does have an effect on music perception.  In particular, younger people tend to jack the bass up to levels that seem drowning to older ears, because for them everything is swimming in treble they won't be able to hear in a decade or two.  Please don't throw a "cite some sources or shut up" asinine remark at me; I'm not proposing a major revision to the article (about which I care little; my watchlist is HUGE). Just suggesting that something is missing, and those who do watch this one could improve it in that direction if they get some time to go digging around for sources. 68.84.153.245 (talk) 18:44, 11 December 2011 (UTC)