Home Propeller Head Plaza

Technical and scientific discussion of amps, cables and other topics.

My Reply (long)

Bill wrote:
"We need only for it to respond *in some audible way* to the changes we are trying to detect.
AND
It only matters that some audible trace of the difference is left in the record/play process."

I tried to explain this in my original post, but I didn't think it would have to be spelled out word for word.

OK, let's look at that example I made, of the soundcard with the last bit missing, everything below -90 dB is GONE.

Now, we record the output of DUT A, and of DUT B using that same soundcard. We play back the DIFF file, and low and behold, we hear nothing but a slight amount of noise. This means that there is no difference between the DUT A and DUT B, right?

Nope, in fact, they could sound quite different, and we would never even know it if all we used was the DIFFmaker test.

As an extreme example, let's say DUT A _ALSO_ lost the last bit of data, everything below -90 dB was GONE. Unit B does not do this, and is quite accurate all the way down to the 16th bit. Obviously, if we listened to DUT A vs. DUT B, we would hear this complete loss of low level information, the truncation of the audio signals. But the soundcard would be inherently unable to record and thus, show a difference, because IT TOO was limited to this level of resolution.

So how good the soundcard is, makes a difference in how well we can detect a difference between two other DUT’s being recorded.

I know that some have been proposing that we examine the difference between the input of DUT A and the output of DUT A, thus eliminating any loss of resolving power between the two units A and B, but the same problem rears it's ugly head, if we tested DUT A using that hypothetical soundcard, and adjusted the gain for the two different output levels, and listened to the resulting difference file, we would again hear virtually nothing but noise, thus we would erroneously conclude that DUT a was "perfect", that is, it had no difference between it's input and output. DUT could also be tested and also found to be "perfect", but in reality, the two would have quite different performance capabilities, yet the DIFFmaker test would not reveal ANY differences between the two DUT's.

This is why I am saying that whatever the limitations are of the soundcard used to make the recordings, this is the limit of the resolving power of the DIFFmaker test.

Now, I know that we are not dealing with soundcards limited to 15 bits of resolution, that was a gross example to make my initial point that the soundcard can and WILL adversely affect the results if it is less than perfect itself.

I once again point to widely accepted opinion and professional practice, that NO soundcard is considered sonically perfect or even sonically SOTA, much less a $150-200 soundcard, therefore, the DIFFmaker test is going to be limited in how much resolving power is availble to record differences.

If we postulate that no soundcard even approaches the SOTA sonic quality as exemplified by such components as Mark Levinson, Krell, Audio Research, etc, then we must also acknowledge that the DIFFmaker test will NOT be able to provide a valid difference file for one of these components, nor will it be able to provide all of the subtle differences present and of interest in audio cable, and many other audio components.

None of this addresses the obvious issue of the gain adjustment when trying to do a DIFF maker test for an input/output test of the same unit, obviously, if the soundcard used does not adjust ONLY gain, but introduces any other signal aberrations when a gain adjustment is made, then this distortion of the soundcard will be added to the resulting difference file.
In the hypothetical case we have been examining, instead of hearing just noise, we might hear some very small residual of distorted music, just enough to look as if the DUT was slightly less than perfect, BUT, with a difference that was very low. This might induce us to believe t6hat the DIFFmaker test had actually been wholly unsuccessful and had provided valid and worthwhile data, even though the entirety of the 16th bit and below was missing.

Bill wrote:
"Just look at where your reasoning leads --- Z"

Indeed, it leads to the inescapable conclusion that the DIFFmaker test is not magic or supernatural, but is limited in the same manner as all other audio systems.

Can we hear 'past' a weak link component? Yes, to a certain degree, but I think that it is wise to point out, that with a RECORDING process taking place, anything lost is gone forever, that once it has been 'left behind', it is never coming back, while with real world audio systems, we can can kinda sorta listen past a weak link component, and when that weak link component is upgraded, we can now hear what the other components are doing that we were not fully able to do belfore. However, if the soundcard is upgraded or replced, all the old recordings are still limited in what they can show or reveal, no new difference information will become known until we make that recording comparison again, using the new soundcard.

Bill wrote:
" Say there are differences between the cases being tested that, if isolated, are below -90 relative to fullscale, and are audible. So do these differences have *no effect* on the recordings made? Of course not. They will change values determined by the converter at many sample points. At some point, perhaps one sample value that had been on the edge of being +12,345 instead of +12,346 will get tipped over to +12,236 due to the tiny "difference" being made. An effect will show up in the subtraction result."

Of course, some effect of a difference down at -93 dBFS will still alter bits higher up the ladder, BUT, these changes are greatly reduced, and would tend to fall, by definition, into that realm where they would register as if they were less than a true difference at -93 dB.
In other words, loss of the entire 16th bit will render the difference virtually deaf to signal aberrations below -90 dBFS.

Bill wrote:
"And (back 'under the top), the actual dither or noise with real hardware actually allows signals to be resolvable below the least significant bit, just as sounds can be discerned below noise from analog sources."

Sure, but now we are limited to 15 bits and a little more with dither, instead of 16 bits and a little bit ore with dither. BTW, dither designed to augment and extend a 16 bit system WILL NOT adequate exercise the 15th bit to "fix" it, the dither level would have to be increased correspondingly, and so, you argument falls apart in a very real sense.

Bill wrote:
"But people get to listen to the result themselves, they don't need me (or you) to tell them what they can or can't hear or what they think may be significant."

Of course, but what they may not know, probably DON'T know, is that the difference file is still limited in a very real sense and manner, to the sonic quality of the soundcard.

It still boils down to whether or not you think that the specs are enough, or the sound as it listens is what matters, if the soundcard is not SOTA sonically perfect, then you can not test SOTA audio components with it and expect to see/hear all that might be there.

Bill wrote:
"And why the concerns about -90dB detections (reduced from the -120dB argument in an earlier post)? "

I think that you must be referring to someone else's post here. My argument is that the known and obvious changes in the sound when different dither algorithms are used, is a form of proof that signal aberrations that are at or below -90 dBFS DO impact the sound. Each dither algorithm does sound different, and not just at low levels, but all across the range of the music signal. Music doesn't 'live' at 0 dBFS or even -10 dBFS, but rather, across a wide gamut of mini ranges as it swells and flows.

Bill wrote:
"(That rules out use of vinyl, you know. An exceptional vinyl recording and setup can maybe get to 73dB S/N.)"

No, no, and no.

First, I have worked extensively with measuring and characterizing the vinyl playback system, I worked for Discwasher for years, and have actually done the research (in some cases, original and unique work). A well pressed virgin vinyl recording can reach a dynamic range in excess of 80 dB, from approx. a clean noise floor of -70 dB (where signals can be heard below for approx. 6-10 dB) in the midrange, up to about +10 to +12 dB above 0 dB (where 0 dB is defined as 3.54 cm/sec modulation velocity per channel, or 5 cm/sec lateral, the Shure tracking test record had levels of 25 and 30 cm/sec, which approximate to +18 dB).

If the record was from a direct to disc session, it could even exceed that range, up to about 85 to 88 dB. Some of the Sheffield direct to disc records regularly hit +18 dB, and had a noise floor of -70 dB or better, all in the midrange.

Second, it does not necessarily rule out the use of vinyl as a test medium, because as we are all aware, you can hear things on a vinyl record that allow you to discern differences between SOTA power amps, preamps, and audio cables. It is not JUST the dynamic range, or the low level resolution, but rather, the entire package of performance during the playback event. Where vinyl excels, is that when it is reproducing the music where it 'lives' at moderate to low levels, it is fairly linear and resolving, without some of the issues that plague digital audio.


Bill wrote:
"Should anyone who can solder want to really hear what 90dB below their listening level sounds like .... "

Wrong. I did not say -90 dB below the music level. Music level implies the average level, not the peaks, and not 0 dBFS. Relative to the average music level, a -90 dBFS signal is really at somewhere around -50 to 55 dB from the average level of the music, which is quite a lot off from the -90 dB straw man you set up.

Bill wrote:
"In DiffMaker tests, though, such masking "other things" are removed in the subtraction. Masking isn't a factor -- at all, unless possibly the noise is so high as to be noticeable. The test is between hearing something or hearing essentially nothing."

Perhaps masking was a poor choice of words on my part, I was not referring to the classic 'masking' as applied to psychoacoustics and MP3's, but rather more in the nature of the poorly performing soundcard occluding or losing the low level signal information to one extent or another, and thus, 'masking' what was going on with regard to the difference file.

Again, in my hypothetical example, the soundcard will literally not record anything below -90 dB, it is MISSING, gone, poof! ALL differences below that level will be utterly and completely 'masked', or as a better choice of words, obscured. The difference file will contain virtually nothing but noise, or perhaps just enough residual of the original signal that it seems as though the test worked, but in reality, YOU COULD NEVER KNOW FOR SURE whether or not the soundcard was adequate to the task, or not, unless you had conducted a huge series of well done and executed controlled listening tests that incorporated numerous controls and verifications, as well as open testing, to try and verify or confirm that the soundcard was sonically transparent.
Without that kind of extremely costly and time consuming verification, you could not automatically assume that the soundcard was god enough, or did not have it's own flaws literally covering up those of the DUT's.

Bill wrote:
"
And there's another one. If I "add something" to A, and add the same something to B, then subtract A from B, that "something" vanishes, per the simplest algebra. Unless the added something can also make the difference we want to isolate VANISH (not change, mind you, but VANISH without a trace), the difference will be left."

Sorry, but the "added something" may not be a simple addition, but rather, an alteration of the signal, a signal aberration would again be a better term. In the case of the hypothetical soundcard which loses the 16th bit, nothing at all is added to the signal, just removed. However, in my extreme example, (in order to clearly make my point), the loss of signal content is a well defined and obvious thing, let's call it truncation to 15 bits. But what about a soundcard that is not so obviously distorting the signal? What if it is 15 1/2 bit s accurate (a not uncommon situation with a 16 bit based card), and that it has a frequency and level dependent loss of signal, these things WILL NOT necessarily cancel out with a set of recordings of DUT A and DUT B, or with an input/output comparison, but will still tend to obscure any low level differences that the DUT's do have that are below the level of soundcard corruption.

Bill wrote:
"Have you tried it already and found no difference in things? "

Actually, I have tried to use a soundcard in a similar manner, by nulling the test signal with the output of a DUT, as well as to compare two DUT's with each other via a difference file.

I manually edited the timing to match the nearest sample, using both 'ways' {creating two diff files}when it was not coincedent with each other (almost all the time).

At first, I too thought that the soundcard should not matter, but found that using a SB16 just would not do the job; even after I tried a studio grade soundcard, the Echo/Event Darla, there was still too much contamination from the soundcard to do the job properly.

I sincerely hope that you do finally understand that the soundcard quality DOES matter, and that it can not be ignored, that it DOES NOT fully cancel out, only a portion of the soundcard distortions are cancelled out. Losses of the signal, non-linear or otherwise, simply do not cancel out. Non time coincident additions do not cancel out (computer EMI/RFI contaminants) either, and their resultant effect on the audio circuitry can be erratic or misleading with regard to how the DUT's AND the soundcard respond to them.

If you can not see this is the case, then I am sorry for you, it is not THAT hard to figure out, especially when it has been spelled out for you.

Jon Risch


This post is made possible by the generous support of people like you and our sponsors:
  Kimber Kable  


Follow Ups Full Thread
Follow Ups
  • My Reply (long) - Jon Risch 16:33:58 03/12/07 (0)


You can not post to an archived thread.