should I dither? (I can't make my mind up (groan...))

And what does the S/N ratio have to do with this anyway? Has that been part of the controversy in some way? There certainly has been no slightest controversy over the fact that the answer to the OP question on whether or not to use dither is that it can’t make any appreciable difference. I just pointed out the best way to assure oneself of that fact.

The effects of dither will only be audible at the bottom end of whatever encoding dynamic range you have available, and in the presence of a loud signal immediately before a quiet one where you might notice the difference, your ears won't be able to percieve it anyway. The more-or-less logarithmic dynamic range of your ears is about +85dB ref .0002 microbar (the normal reference), and above this level, the attenuator built into your outer ear comes into play progressively (which is how you can hear louder sounds without pain), but this has the effect of lowering the normal threshold below audibility temporarily. So the S/N ratio of what you are listening can play a very significant dynamic part in terms of what's actually perceived. The effects of dither are, after all, only heard in signals which have content around the least significant bits - and even if these are still there, the presence of lounder signals will mask them.

Assurances, though, are unncessary, and since the ones you are talking about are both subjective and uncontrolled, they are not of any use anyway; they would be of no more use than any other psychological crutch. The simple physical facts are all you need to know, and they are well established. And what they amount to is that with 16 bit material having a reasonable dynamic range, you are certainly able to percieve the difference between dither and no dither. With material having a very narrow dynamic range, played loud, that lasts for more than a couple of minutes the chances of you being able to detect anything at all at the 0dB point are pretty slender, never mind being able to work out what you heard!

So the “fact” that people think they hear all sorts of things that are not supported by the physical properties of the particular sounds is itself an illusion and ABX tests provide no evidence about what people “actually” hear, and there is no possibility of you ever being confused about what you hear because you know the real facts, and those facts don’t influence your expectations or beliefs, they just auto tune your hearing into being absolutely correct?

How could I argue with that?

So the “fact” that people think they hear all sorts of things that are not supported by the physical properties of the particular sounds is itself an illusion and ABX tests provide no evidence about what people “actually” hear, and there is no possibility of you ever being confused about what you hear because you know the real facts, and those facts don’t influence your expectations or beliefs, they just auto tune your hearing into being absolutely correct?

How could I argue with that?

That's quite correct, and you can't. The moment you said 'think you hear' you established the total  lack of credulity of this so-called 'method'.

http://audiomastersforum.net/synforum/0/topic-802.htm

This just get more and more weird Steve. People think they hear things under sighted conditions, such as differences between cable A and cable B, or difference between dither vs no dither, that they can’t hear under blind conditions. Or during blind testing they still think they hear whatever they thought they heard under sighted conditions, but the test reveals they did not actually hear it.

Then they either accept the test logic and stop hearing what they originally thought they heard, or they react against the test concept and come up with justifications why such test can’t possibly be any good, and refuse to participate any more. Many people are too intimidated by the threat to their beliefs to even try the first time.

Furthermore, under  conditions that are not double blind, i.e. the subject can receive cues from test controllers or from environmental conditions other than the sound itself, they can be made to fully believe they are hearing something they are not. That is, they can be deliberately manipulated or misdirected to think they hear something other than what they actually hear.

They can listen to test sounds under double blind conditions and correctly report what they hear. Then on subsequent trials, they can be deceived into thinking they will hear something other than what is actually presented, using the same test sounds that they originally heard correctly. They then think they hear what they were led to believe they should hear.

But surely you know this. Do you believe that the basic technique of double blind testing is not credible in all disciplines where it is used, or only on the occasions where I employ it with audio?

Do you believe that the basic technique of double blind testing is not credible in all disciplines where it is used, or only on the occasions where I employ it with audio?

Belief doesn't come into it, and not it's not about you. The so-called 'test' (it's not really a test, but a poorly controlled experiment) is not credible at all under any circumstances. I could tell you why (and the reasons are good) but I don't have time to go through all the details at present. Briefly, even if you could get rid of all the room problems (by locking your head in one place) by using earphones, you'd still have all of the perceptual problems to contend with, and those you can't fix at all.

I suspect a lot of researchers would be dismayed to learn those reasons, although your reply seem to ignore the question of the method’s usefulness when applied to fields other than audio, as it is, with great frequency. It seems a bit far fetched to say the technique is credible when used in these other disciplines, but not with audio.

Surely you are not advancing a claim that these room and perceptual problems you  mentioned won’t be the same problems in a sighted comparison? That seems to come down to “use your ears” being totally worthless advice under any conditions. Neither sighted nor blind comparisons are credible.

Since it is audio perception that we are trying to get a handle on anyway, saying that problems with perception prevent us from getting at anything useful seems rather nihilist. My position, hardly unique, is that sighted comparisons involving threshold or subtle differences are not credible, double blind comparisons are necessary. I’ve already written here about the reasons for that viewpoint, although I don’t say I’ve covered everything.

From your position, ow could we get any information about perception, at least about sound perception? By only accepting measures we can make with instruments that don’t depend on human consciousness involvement to get their readings? How would we know what meaning or relevance to ascribe to those readings if it is useless to try to correlate them to reported human experiences? And wouldn’t those room and perceptual problems still influence many strictly physical measurements, perhaps depending somewhat on how “perception” is defined?

I suspect a lot of researchers would be dismayed to learn those reasons, although your reply seem to ignore the question of the method’s usefulness when applied to fields other than audio, as it is, with great frequency. It seems a bit far fetched to say the technique is credible when used in these other disciplines, but not with audio.

Since we are talking about different senses here, I'd say that your opinion is the far-fetched one. And yes I'm only talking about audio.

Since it is audio perception that we are trying to get a handle on anyway, saying that problems with perception prevent us from getting at anything useful seems rather nihilist. My position, hardly unique, is that sighted comparisons involving threshold or subtle differences are not credible, double blind comparisons are necessary. I’ve already written here about the reasons for that viewpoint, although I don’t say I’ve covered everything.

Just because it looks like one interpretation of nihilism doesn't mean that it's wrong. You should look at all of the research done into audio perception (of which there is a long thread on this forum, started by Ozpeter) which basically rubbishes, by careful research, a lot of what is claimed about audio perception.

From your position, ow could we get any information about perception, at least about sound perception? By only accepting measures we can make with instruments that don’t depend on human consciousness involvement to get their readings? How would we know what meaning or relevance to ascribe to those readings if it is useless to try to correlate them to reported human experiences? And wouldn’t those room and perceptual problems still influence many strictly physical measurements, perhaps depending somewhat on how “perception” is defined?

The idea of 'perception' in this context is that we are listening to a performance, and if you are listening to it properly, then you will not be listening to the sound. In other words, it has to be content-based. And even if you are listening only to the sound (which in the presence of content, will be rather difficult), your ability to detect even large and measurable differences by ear even over a few seconds is minimal - and if you listen over longer periods, it's pretty much non-existent. And when you take into account the differences in peoples' state of minds on different days, and things like different tiredness levels, the whole thing becomes utterly pointless, simply because you can't normalize any of that. What you like on one day may be completely what you don't  like the following one.

We established many years ago what aspects of audio performance made an impact on peoples' 'ease of listening', and guess what? We did it without any reference to ABX at all. With audio, the whole ABX thing is a giant red herring.

And on top of all that, you're not being very consistent, are you?

On the one hand, you are advocating a system here that performs a so-called test, although there is no way of objectively analysing any results because you haven't measured anything - and on the other hand, you are desperate to measure a meaningless number when aligning a cartridge.

I think it's about time you got your ideas sorted out a bit. But there again, I've thought that for quite a while...

As far as I'm concerned, you can measure the results of tests on audio equipment - and we pretty much know what factors affect what audible results. But when you listen to the material, you are not supposed to be 'measuring' anything; this is purely a subjective experience, along with all the uncertainties it brings. If you are recording live material via microphones, then you have things you can do to alter the results, and these make far more difference than any minute distinctions that you thought you might have heard during a so-called ABX test. ABX tests have no place in any type of audio assessment whatsoever.

AndyH, nothing I posted was intentionally an ad-hoc attack on you.  Which is not to say I was not being critical of what you posted.  You entered this thread with this statement:

It need not be subjective.

As a direct counter (which if I were unduly sensitive I could easily construe defamatory, though whether libel or slander will still require successive legal challenges and some time to see how digital publication permeates the culture as it arcs towards technical maturity) to an assertion I made that, with regard to OP's query, the choice about whether to dither or not was subjective.  You then posed something (that I will for convenience continue to refer to as a 'test') that is not only categorically subjective.  But fails to be able to resolve the issue in any case.

You followed up the initial post with others that demonstrated a confused understanding of some underlying principles regarding digital audio. For example you seem to have conflated quantization error with distortion.  Yet even if you are using 'distortion' in vernacular, to mean anything that deviates from an arbitrary source to perceive quantization error and distortion as congruent is not particularly useful.  It tends to suggest that you are still not particularly clear about what goes on with ADC, DSP, DAC.  Additionally your stated belief

That is, besides the fact that there is so much broadband noise coming off any LP that, even though it isn't ideal dither, it alone is adequate to assure there is no quantization distortion.

reinforces the opinion that quantization error and distortion are equivalent and that both can be banished with a single stroke. Which is not true. Dither can not 'banish' quantization error at all.  Additionally suggesting 'broadband noise' in the source obviates the need (if there is one) for 'dither' to mask quantization error when converting from a larger to smaller bit depth suggests in turn that you probably do not understand 'dither', might not be conversational about digital audio.  Even on general purpose Audio forum it should not be necessary to spend pages detailing the limitations these various claims present.

While the above, brief, outline is not intended to be defamatory, it is meant to suggest why a critical response to your posts is not untoward.  Admittedly there are times, over the years, where I am genuinely not sure whether you are 'putting people on', 'jerking the chain', so to speak to trigger some hyperbolic response.  As in:

I don't dispute that. Sometimes I welcome a bit of conflict to spice up the dialogue,

. . .

Or are fundamentally confused concerning discussions into which you insert yourself.  But no matter which, there are times when a critical response is suggested because letting disinformation stand can, at times, be problematic down the road.  That is, you repeat something you've read, insert, out of context, into a query, then at some future point someone quotes you as being a definitive response to something else.  Which is one of the issues the SOS oped piece, which triggered in turn Steve's criticism of SOS, was complaining about. 

A 'test' whose results are limited by the object, can be considered to be an objective test.  One in which the results are limited by the subject, is of necessity a subjective test.  It is almost ridiculously easy to establish an objective test that confirms that a 32 bit(float) digital file that has been dithered, then converted to 24 bit(integer) is significantly different from the same file converted to 24 bit without the dither.  As there is no functional objective challenge to this the need to perform this process is subjective.  Dither, in this case, can be demonstrated to do what it is intended to do: mask the artifacts of quantization error, in the step down.  It should be obvious that results of an individual ABX 'test' are limited by subjective criteria.  This fact in turn deals with your opening salvo in this thread.  You have not introduced a non subjective test in support of:

It need not be subjective.

It is possible to construct a black box experiment in which subjective response is subverted, subsumed, to produce statistically significant results.  The experiments are notoriously difficult to set up, results are universally (if anyone reads them) open to peer challenge and lie at the heart of ad-hoc, defamatory, slanderous, vicious debate between so called 'hard' and 'soft' science.  Between those who stick electrodes into subject brains and those that talk to subjects.  Sound is amenable to objective tests.  Hearing is a far more difficult subject.  An ABX 'test' , limited to a single subject, where tester and testee are the same, administered in an uncalibrated environment, which does not address, let alone limit, even a fraction of the non linear variables associated with human perception simply fails to meet minimum criteria for any statistically significant test. 

Or as P. Simon said, '. . . A man hears what he wants to hear and disregards the rest.' Followed by a bunch of dithering lie, li lie da lie lie lie . . . Sound and hearing are not equivalent and no ABX 'test' is far enough from 7th Ave to be statistically significant.

So while the 'need not be subjective' challenge is dispatched pretty much out of the gate it might leave open the question as to whether an ABX 'test' might be subjectively useful for resolving the OP's query.  And as AndyH has so graciously, if with a certain amount of displacement, pointed out only the OP can, subjectively, make that determination.  But it is a question that can be examined and some of probabilities as to effectiveness can be addressed.  I do not remember for sure but think AndyH has indicated that his DAC hardware included an Echo MIA and Audiophile 2496.  While AndyH does not seem to believe that S/N ratio has any bearing on the dither issue:

And what does the S/N ratio have to do with this anyway? Has that been part of the controversy in some way?

Let's pretend for a minute that it, S/N, does bear on the issue.  If Steve can be believed, (and I'll go way out on a limb and suggest that, in this case, at least, he can be (for the ironically challenged this was strained sarcasm)) then any audible artifacts of dither will be 'located' at the noise floor, if noise shaping is employed we will try to push the artifacts into frequencies at the upper limit of human hearing. (That, for this debate, is in turn significant because of psycho-acoustic human response to those frequencies and the energy required for human perception of them).  If dither was not associated with 'least significant bits' it is doubtful that it would be an effective tool for masking quantization error. (so this would not be a query at all)  An Echo Mia has an effective DAC of something a shade under 18 bits.  The Audiophile is spec'd a bit worse.  (With Echo tending, in my experience, to be conservatively spec'd while M-Audio is a bit more optimistic.)  In an case, if I remember correctly, AndyH's hardware supports between 17-18 bit DAC.  No matter what one chooses to believe about ABX 'tests' in general, for this specific instance, based on specs that are consistent with majority of entry level prosumer gear, the hardware DAC simply lacks the resolution for hearing to distinguish between a dithered and un-dithered 32-24bit conversion, regardless of any other variables.  ABX 'test', questionable (with kindest interpretation) in theory; simply useless, with regard to OP's query, using any readily available DAC hardware.

The same fundamentals, not esoteric formula secreted and passionately defended by self proclaimed experts (a title I'd reject, for myself, vehemently under any circumstance, I'm not merely one of the blind, eschewing any leading of other blind, but am a 'story teller', the scope of my invention (here stories, music, mixes are more or less synonymous) pretty much excludes me from guild of 'expert' under any circumstance), that exclude an ABX 'test' as tool for resolving the OP's query also suggest why such a test is unnecessary.  If you understand any of the relationships among digital files that encode for sound and human perception of the audio produced upon DAC then you pretty much know all you need to know about whether dither is significant in the OP's query.

(I didn't spend any time for critical review of Younglove's archived thread.  But I am not sure that I would categorically disagree with his assertion that en route to 16bit, if you pass through 24bit that dither between 32 & 24 is useful.  As I started editing, processing, mixing @ 32bit, more or less exclusively, in '96, a time when mastering facilities might lack ability to deal with floating point files, this was not entirely a moot point.  Generally speaking I was a proponent of the conventional wisdom, 'don't dither the dither' . . . But?

Even now I am at times torn when I am forced to migrate a project into the PT environment about whether to dither or not.  In most cases, particularly when I will retain control of the mix, I don't . . . But, subjectively, as I said, I remain 'torn'.  Don't lose sleep over it of course.  Just one more variable of the thousand decisions any mix entails.)

To recap: whether to dither or not (with regard to OP's query) is purely a subjective choice.  There is nothing to substantiate that it will have any audible influence at all.

(now have to decide whether to come back in another six days and see if debate continues . . .  it's not that I don't have life!)

Even now I am at times torn when I am forced to migrate a project into the PT environment about whether to dither or not.  In most cases, particularly when I will retain control of the mix, I don't . . . But, subjectively, as I said, I remain 'torn'.  Don't lose sleep over it of course.  Just one more variable of the thousand decisions any mix entails.)

I recall reading somewhere that there's something screwed about the way that PT handles dither when reducing bit depths. I don't recall the details precisely, but I do seem to recall that you have to do the dither part of the conversion at a stage that you certainly wouldn't consider correct in any other software... bottom line was that if you send a 24-bit int file to PT, don't dither it, but leave that almighty pile of pooh to make, and then sort out, the mess itself.

I recall reading somewhere that there's something screwed about the way that PT handles dither when reducing bit depths.

A few years ago people would have to add a plug-in with internal dither to the signal chain before changing bit depth (or even changing level) in Protools. I know that they later made a dithered mixer as an alternative to the standard mixer but I don't know what the current situation is.

James.

SteveG
An ABX test is measuring whether or not a person can correctly identifies which signal is which. The measure is the number of correct and incorrect responses. In the use of ABX software programs, a device not easily subject to subjective influences (the computer system) is collecting the data, then making the measurement . It doesn’t have to be on a computer system but then it is general harder to do and definitely harder to control. The analysis is statistical. So is the analysis in many areas of “hard” science.

The more common purpose of an ABX test is not to listen to the performance but to listen for differences between whatever signals are being compared. What one likes or don’t like about the audio isn’t generally relevant, although that aspect could be used as an operational criteria, but more easily with just an A/B comparison than an ABX test. Identical copies of a performance are used for both signals of the test, if music is involved.

Data from hard measurements, or any other source of information about the subject to be investigated with ABX test, is at most a side issue. What a person should hear, or should not hear, is not the object of such a test.

If a person is convinced, from sighted comparisons, that there is a marked difference between the source and an mp3 version, but cannot distinguish between them in a blind comparison, what do you conclude? If it only happens one or twice it may indeed be meaningless, but if it is consistent, something real is involved. Is there another believable explanation?


By what value system is the tracking force measure meaningless? If it were possible to know that the recommended maximum had been exceeded by some aspect of the signal coming out of the cartridge (other than something really extreme such as no signal or some nasty noise because something just broke), that might be fine, but I’ve never seen anything about such information.

I have read reports of tracking, on test tracks, continuing to improve well past the specified maximum tracking weight. That was what I seemed to be getting. But without a decent measurement of the tracking force itself, the cartridge signal is just an observation on a gradient without any reference points. IN the end I had to determine the answer myself.

oretez,

and the choice is thoroughly, completely, aesthetically, magically subjective

Basis grammar tells us the subject of your statement. I can’t come up with any logical reason for someone to assume that my statement has a different subject, such as “the dither” or “the distortion.”  or “the efficacy of dither.” It is very clearly about “the choice.”

Nothing I can find in my expounding about my declarative statement says any thing about hearing not being a subjective experience. I pointed out a way to gather objective evidence about one’s subjective experience. If it isn’t clear already, the evidence is the correctness, or incorrectness, of the subject’s response when trying to differentiate between the two signals being compared.

This evidence allows the individual who takes the test to make that choice based on something other than their subjective experience so ... it (the choice) doesn’t need to be subjectively based.

I did not suggest that the tests will reveal any basic truths about physics or biology. The test results only applies to what the tests are properly used for. That isn’t to gather information on basic physical processes. It isn’t about what might happen with a different listener or different equipment or while the subject is on some exotic new drug -- unless enough tests are run incorporating those other variables in an adequately controlled way to fit a good statistical model. These limiting considerations do not make ABX testing unsuitable or unuseful for a task such as the one under current discussion.

It should therefore be clear that the results of such an ABX test apply only under the particular test conditions used. I don’t know where I suggested anything contrary.

A purely subjective evaluation on the kinds of comparisons for which ABX testing is reasonably applicable is not more readily generalized than an ABX test; it has the same limitations as to person, equipment, conditions, etc. People believe unreasonable things about what they listen to even when all the physical aspects deny their conclusions. Much of the audiophile trade would not exist otherwise. The suggestions in this thread that the physical facts alone should be enough for anyone to come to the correct conclusion might be true in some ideal world but clearly often don’t apply in this one.

Effect is a word, like many, with more than one definition. If you can’t accept that my “effectively” is the same as “in effect” is the same as “to all practical purposes,” as I’ve already written on in considerable detail, then perhaps you will persist in your view that I don’t understand the very simple concepts of quantization error or understand what dither does to ameliorate it (make it of no practical effect) in those circumstances where it might actually be of any audible consequence. Hereafter I’ll try to use your term, artifact rather than distortion, for those funny noises that result when the error is correlated with a low level signal.

It may be that SteveG’s example of using dither to repair old recordings that have a fault in reverb tails is an example of simply masking the artifact, but that is not the general operation of dither. The artifact is removed, eliminated, decorrelated, whatever term seems like the most comfortable way to express that it is no longer present. In audio in general, one signal masking another does not mean that the masked signal isn’t there, only that the human auditory system can’t detect it. A spectral view still shows the masked part. That isn’t true with quantization error artifacts and dither.

It is the dither noise itself, not the no longer present artifact, that is shifted towards the less audible frequency range by noise shaping.


It might be interesting, however, to find out why you think the broadband noise in an LP recording isn’t “effective” dither.

LP broadband noise may not all be truly random but it generally does not correlate much with the audio. It is almost always at a significantly higher signal level than necessary for effective dither. Not only does it seem very logical (to me) that the broadband noise will decorrelate the quantization errors from the audio, thus (effectively) eliminating the distortion, sorry, artifacts, but my experiments seem to substantiate that completely. This is not just based on my subjective experiences.

I allow argument that there can be instances when some LP noise clearly does vary with the music is an annoying way so that its effectiveness as dither is likely significantly diminished, but this is not the usual case.

[PURE BS.]

SteveG
An ABX test is measuring whether or not a person can correctly identifies which signal is which. The measure is the number of correct and incorrect responses.

That's not 'measuring' anything, that's simply counting. There are lies, damned lies, and then there are statistics. Instead of blindly parroting somebody else's misplaced ideas of whether ABX has any relevance to audio, why don't you actually try thinking about it for yourself instead? And if any result from an ABX audio test shows a repeatable difference, then that difference is going to be distinguishable anyway, just from normal listening.

The more common purpose of an ABX test is to listen to the performance but to listen for differences between whatever signals are being compared.

I feed those two outputs from whatever, or signals, into a single converter, and subtract them from each other. That way I can actually determine what the difference is - I don't need a subjective observation system for that, because it doesn't reveal any useful results.

Data from hard measurements, or any other source of information about the subject to be investigated with ABX test, is at most a side issue. What a person should hear, or should not hear, is not the object of such a test.

That's PURE BS. It is a completely meaningless statement. 'Should hear'? Get a grip on yourself.

By what value system is the tracking force measure meaningless?

I have read reports of tracking, on test tracks, continuing to improve well past the specified maximum tracking weight. That was what I seemed to be getting. But without a decent measurement of the tracking force itself, the cartridge signal is just an observation on a gradient without any reference points. IN the end I had to determine the answer myself.

I think you just answered your own question...

It may be that SteveG’s example of using dither to repair old recordings that have a fault in reverb tails is an example of simply masking the artifact, but that is not the general operation of dither. The artifact is removed, eliminated, decorrelated, whatever term seems like the most comfortable way to express that it is no longer present. In audio in general, one signal masking another does not mean that the masked signal isn’t there, only that the human auditory system can’t detect it. A spectral view still shows the masked part. That isn’t true with quantization error artifacts and dither.

It is the dither noise itself, not the no longer present artifact, that is shifted towards the less audible frequency range by noise shaping.

I think that you should stop pontificating about things that you clearly don't fully understand.

I'll try to make it very clear about the dither thing, because it's really quite simple. If you add noise to a signal after it has been recorded or converted to 16 bits, all you are doing is masking the quantisation error. To dither a signal correctly, you add the noise to the signal at the correct level before the conversion. It is the statistical manipulation of this pre-added noise that creates the dithering effect, and the increase in resolution. The basic techniques only differ in their spectral distribution of added noise, but the more advanced techniques use the signal itself to determine what should be added.

You don't just have to take my word for it. Here's the relevant bit from the Wiki article on dither:

"Dither should be added to any low-amplitude or highly-periodic signal before any quantization or re-quantization process, in order to de-correlate the quantization noise with the input signal and to prevent non-linear behavior (distortion); the lesser the bit depth, the greater the dither must be. The results of the process still yield distortion, but the distortion is of a random nature so its result is effectively noise. Any bit-reduction process should add dither to the waveform before the reduction is performed."