Right, time to report a bit about this.
[Edit : I removed the pictures from this post, while the text was not changed accordingly]
First of all, the measurement equipment (officially there for these kind of things) brought nothing. I guess by now this a kind of known, partly (or mainly ?) because I ranted about it a bit elsewhere, and I seem to have one follower at least.
So, for weeks I have been thinking how to approach this, and while working on the "glitch" software two weeks back, I suddenly started to see some lights.
Two weeks later, which would be last Friday, the first part of a program to do it was ready, and this weekend I have been measuring and measuring and measuring. A brief view in the directory I use for it, shows that I took 133 recordings of an 80 second track, and below Compare01 may give an impression of what I compared.
The main conclusion is :
oh yes, I can measure the difference !The means used comes down to recording the music data from the analogue out from the DAC, and compare those recordings amongst eachother.
As you can see in Compare01 for each same version two recordings are taken, which is necessary for the interpretation of the correctness of the recording itself. More difficult to explain (so I won't even start that) but it is crazy difficult to obtain compareable recordings, which is related to offsets, sample rate and clock alignment(s). This summarized, it took me days already to get similar results from the same situations.
On a side note, keep in mind that for each difference measured, the bit perfectness is still obtained. This now can be checked by means of the same software very easily, with the setup of looping back the digital line into a digital in of the soundcard. This by itself is no exciting thing, but the fact that the same software is just, showing zero differences always, is a nice thing to know for good feelings (about some decent wat of working).
What do you see when the same settings are used ?For XXHighEnd this would mean the same Q settings and all, and in general this is about comparing a player with itself, performing two runs with the same settings.
First let's see what's in that 80 seconds 16/44.1 file :
There are 80 x 44100 = 1,764,000 samples for one channel. Note all measurements are about one channel, which just is sufficient (for this purpose and for now).
To my findings so far, when a random *same* situation is compared, around 82% of the samples will be equal.
This means that for our example around 317520 samples are not equal.
Keep in mind that the maximum value in one sample (for one channel, 16 bits) is 65536, but that any situation which is officially equal will have a maximum off value in any sample of 1 (and more rarely 2). This gives an off value of 317520.
See below Compare02 for an example.
These ratios apply always and spring from general phenomena not under our control,
but count for my DAC in this case. It will depend on things I currently don't know, but I have a most strong feeling somehow the clock of the DAC is influenced. I derive this from everything I saw at comparing and the patterns I saw. There is more to it though like the receiver chip, and of course jitter coming from the SPDIF connection I used.
If you take jitter as the example of something not being under our control (once the DAC is there), it is an explanation of not every run at comparing the exact same settings, ending up with the same figures. So keep in mind : at running the same settings, only 82% of samples will be the same, and the off value in one sample will be 1 only. I assume this is inaudible when you run a track twice, of which I just have proven it will not sound the same, because it just doesn't come out of the DAC the same.
What do you see when the settings are different ?Simple : the percentage of samples being the same goes down, and the off value will rise. But there is a very profound other difference :
The off value per sample will be higher than 1 for many samples.See below Compare03 for an example.
I'll spare you the longer list for now, but this is from a relatively mild difference in settings, and the percentage of samples the same dropped from 82% to 81% only, but the off value rased from 317520 to 361724.
This is a mild difference, but something more bad is going on : it happens in patterns. So what you see in Compare02 recurs throughout, and no matter how small the increase of the value per sample is, it recurs every few ms and it will imply an audble frequency by itself (think about dither and why it should be random and you'll understand).
Compare04 below shows a much more severe case;
Here the number of samples equal dropped to 40%, and the total off value rose to 1,980,277. This is not only because the spikes are deeper, but merely because of a more consistent off value, which you can see in the screenshot (the parts without spikes, but with a consistent +++ etc.).
Keep in mind : this is bit perfect.
I have seen cases comparable to Compare03 but worse, which are beyond my capability of explaining so far. There's rows and rows and rows with the value just being of in a rather consistent fashion. It seems that something can get confused and won't recover for a while. But, it always does ... until it happens again.
Not to forget : when I'm talking about things getting confused, this has to be seen in the context of the phenomenon happening, can be repeated over and over. With this I mean, that the specific comparison of two (settings) situations showing this, do not show when the two files recorded from that same setting causing it are compared. Then it's just the known 82% of samples being the same, the others being off by a value of 1 (and rarely 2).
So whatever it is causing this, just can be repeated. Of course this is just the Q settings, or the player for that matter.
Further explanationI explained the above without all the context needed, thinking it would be more easy to grasp. But what actually is happening, is this :
If you look at the screenshots again (03 and 04) ... you see the
relative difference between two recordings. Or in other words, you're not seeing the difference between the original source and the recording. It all is crazy difficult already, and that is just another step for me. Besides that, comparing with the original recording would be impossible with an oversampling DAC which changes the data in the first place (I tried it, around 1% of equal samples are left then). Comparing recordings each from the oversampling DAC, however, just works
So it is about relative differences ...
This means that the pattern I talked about, and that by itself turning into an audible frequency, may indeed be so, when the setting I used for the base coincidentally is the absolute correct one. And I don't know that ...
Keep in mind : whaver setting I use and take two recordings from, show the same "being good".
What it comes down to, is comparing many settings, and learn from the relative differences in between those. If A-B shows a pattern, but A-C shows another pattern which also shows at B-C, you can bet C is creating that pattern. Once this is known, it is a reference by itself.
From this kind of operating I learned that a very high Q1 value compares very much to a very low Q1 value, which btw was the very last I expected.
To end this for now : Foobar WASAPI shows a mild pattern compared with XXHighEnd WASAPI (Engine#3) just the same, when compared with XX-Q-4-0-0-0-0 (which seems safe). So far, I don't know whether Foobar creates that pattern, or XX does it. We only know XX sounds better, but we actually don't know whether that's because an unauthorised frequency riding on things. So now it becomes more complicated, and the Q sliders can be used to simulate something which is going on in Foobar. When the comparison shows the normal 82% and off value, we'd be listening to Foobar ...
But it's not quite finished.
Along with some other analysis stuff, for your benefit, it will be in 0.9y (will take a while !), although I guess some of it will be available in the licensed version only. Must see about that.
Peter
