[Sorry for a long post, probably full of typos ...]
I've just been listening to Keith Jarrett's Koln Concert. Wow!
I've never heard the ambience of the concert hall like this before. And for me, this is the key difference between 'Arc Prediction' and 'no upsampling'. The latter gives the impression that everything has been recorded in an anechoic chamber. With the former however, you can hear the natural reverberation and decay of the hall. Stunning!
So, where exactly has this extra information come from???
I guess this needs one of those 3 "papers" I have written, and I guess some day I will have a nicely worked out (and understandable) version. But generally it comes down to so many things being wrong. Here's an excerpt of one of the versions of the paper. This starts with a picture (not shown now), and explains about how wrong things actually are :
This is just the analogue wave as it will be output (from the player) to the DAC. But hey, this looks nothing like a sine ! And indeed it does not, and this is because when the samples were taken (think from left to right on the x-axis = the time domain) the sample points did not meet the peaks of the sine, but instead they were taken somewhere at the slope of the sine. You can see that occasionally the peaks were taken (look at the maximum values), but this is more often not the case.
Now, try to imagine that the max peaks as you see them are at full scale, and that those you see just a little under already are twice as soft. Notice that yo must think "dB" here, and a little under as you see it here is approx. 6dB, or twice as soft. If you look in the area of sample number 79 you can see that only half of the normal full scale is reached, or IOW this is at -48dB. This will give as good as no sound to your ears.
If we think decimal, that -48dB is at half of the full scale which is half of 32768 (for either plus or minus side), which means an error of decimal 16384 on to 32768 which of course is 50%.
This is the job : Turn that error of 16384 into an error of 0, and see that there are 16384 possible steps of improvement !
And so you see, while we may have thought digital is quite OK, it is actually the most wrong, with the too low sample rate as the cause. And thus what (software) engineers strive for, is getting that analogue wave back to normal.
So what I am telling here, is that everything output to the DAC is so hugely wrong, that, well, it can be hugely improved. Of course, this is a normalish task of the DAC itself, but the means used (technical : sinx function) is even more wrong ...
... And it is here where the paper(s) start to end in needing much more attention at laying it all out (with proper graphing and comparisons), and where they thus all end, for now.
As I am always telling that so many things are wrong, likewise I am telling for a longer time now that official measuring and measurements is/are wrong. This in very brief now : it is no big deal to show a 0.0001 % harmonic distortion on a continuous single frequency (or even two frequency) wave created by the sincx function (call it "filtering" the DAC applies), just because it is such a continues same wave, while sincx loves that. It works by "averaging" (*very* roughly said !), and averaging an ever same wave will do no harm at all -> 1+1+1+1+1+1 / 6 stays 1 = the original. But what if it's 3 + 1 + 7 + 2 + 1 + 2 / 6 and this 2.67 not being equal to the original(s). The average is correct allright, but each single input (meant to be output !) has been destroyed. Now, the first case is measurement, the second case is playing music, and sadly that can't be measured.
What it (again in brief) comes down to, is that NOS is way wrong, because that would measure (single frequency) something like 1+1+1+1+1+1 / 6 must be 1.2. And since 1.0 is 20% off to the expected 1.2, this is 20% harmonic distortion. This is measured like that, and it *is* true. What I do, however, is something like 1.4 + 1.1 + 1 + 1.3 + 1 + 1.4 / 6 = 1.2 and now it is conform the expectation, and thus 0.00001 (or 0.0 in this example) % HD.
The big difference with normal "filtering" is that I do this on a per sample base, while sincx filtering can only do it with a (theoretical) infinit number of samples. But say 100 samples, and get the grasp of now one 100 samples all being of on *another* way. It measured good for theories, but all the samples are all over the place, and actually this is more off than "just NOS" (and it is this "more off" for me to proove by numbers).
Back to the actual question "where exactly has this extra information come from???", I could say "so very wrong it was !". And actually this *is* the answer. But, the stupid thing is that this indeed brings a kind of information which is completely new to us audiophools. I must say, at working at the DAC similar things happened to me, but this is still in a quite different leage;
On a side note, and this is important to ever get there I thinnk, keep in mind that my ears seem to be good enough to know that NOS/Filterless sounds better than any type of OS/Filtered, no matter how hugely bas NOS/Filterless measures. And might you not know, this plainly comes down to even 40 THD+N at certain frequencies (this starts at just under 5KHz, and which is completely logical once you're into this all a bit). So, I sat down to *proove* why NOS still sounds better, and it is all about this. So :
In one of the papers I wrote more extensively about the types of smearing. With my brief explanation above, and the "averaging" type of filtering, it may come as obvious to you that this is an explicit smearing type. It is nothing less than spreading the information of 100 samples into the one sample currently output. Literal smearing that is. However, NOS/Filterless smears just the same, although the type of smearing is a completely different one. It "smears" the one original analoge piece of a wave (expressed in a snapshot sample) by its own harmonic distortion because the samples are at the wrong place. So, harmonic distortion smears, and looking at the longer period of the wave, a clean sine is captured (on the redbook CD) as not a clean sine at all. This is of course what we knew already, but it is the imagination of smearing which is important here;
Without creating 8 pages again, the net result of what Arc Prediction does is that "dryness" as I describe it myself. It is one of the first noticeable things, and dryness can also be described as shortness, the best perceived at bongo like instruments, because their attack is so short (in synthesizer terms : there's hardly an envelope, which is the shape of the attack).
While it is nice to perceive a bongo as a bongo, all becomes *really* apparent when you found yourself disturbed by one of the most uses instruments : a snare drum, and not being able to map your mind onto the snare drum used. I mean, if you have a drum kit yourself (like I bought one 2 years ago, just to compare reality with music reproduction), it is just that instrument that makes you wonder what's actually wrong, because it seems so simple. Today, and relative to how it was, you can just hear what was wrong : the hit on the skin smeared the snare. So, the hit on the skin (even more firmly spanned than a bongo) is now as short as reality, which a. gives it its color, and b. makes the snare sound as how it really is.
Once you got the above (I mean, listen whether I could be right), it becomes more easy to see how everything got smeared, and how the accidental hit of the rim of a tom now allows to be all over, while before it was smeared into the (hit on the skin of the) tom itself. Keep in mind, a hit on a rim is one of the shortest sounds available (attack is as short as can be, and decay is also as short as can be), and today it just has died out before the skin is it. Nothing less than reality, and caused by the shortness/dryness which now is "available".
Again once you perceive this indeed, you can go further in the "stage sounds". So, indeed it is true, and not by a small amount ! There's voices, placements of microphones, walking, laughing, but just the same there's the now all over hit of the (spanish) guitar cabinet during playing the strings. ALL sounds different just because these sounds now are there.
In the end you could say that all now sounds short enough not to get smeared into the nest sound.
The recap is simple :
a. no HD in the audible band;
b. no ringing.
Ad b.
I am sure this is the most important element of good music reproduction. All is a tradeoff between this kind of distortion and the harmonic distortion of NOS/Filterless. And (still assuming good ears), while NOS/Filterless takes prevalence over OS/Filtered, at the first measuaring near 40% distortion, we really may wonder how OS/Filterless measures once it can be done as should.
Lastly for now on this subject, and a real technical matter : Those who are really into this (signal processing) may state that reconstructing the analogue wave on a sample per sample base is not possible. Although at this moment I must see the first one telling *why* that is (which I by now can), the net result is very different, and explaining *that* requires a small book by itself. For later.
Also Peter, when are you treating us to a 'real' CD ripper? When you do, I might seriously give up on hi-res material and stick to 16/44.1 for a while...
As said elsewhere, 0.9y was created for just this. Thus far, however, I just didn't get to it, although the most of the base software already has been written (this must have been April 2009). With ripping as the base, and which
might come down to merits of great importance again, I guess I fell from the one thing into the other, which -you may recall- started with measuring. And, this (detail of) measuring is obviously needed, while normal measurement means won't allow showing the differences (while we sure hear them !). So, to make this small subject consistent with the main subject here, let me tell this, which by now is completely consistent with the knowledge from today, which knowledge (on my side) sure was *not* there back at the time :
The measurement means I created (remember, a.o. needed to judge good/bad ripping) worked for NOS/Filterless. I could judge the data as in the file onto the result coming from the DAC. However, trying this with an OS DAC, and there was no head and tail to anything. Thus, the beautiful graphs I made (actually in XXHighEnd, but all disabled to you), showed completely unrecognizeable waves cmpared to the input waves. It is my luck that I first tried it with NOS, because that proved it could be done. But, and as a side note, already that shows how *again* everything is way way off to what's supposed to be output (which should be 1:1 without any filtering means). Anyway, the waves coming from OS could just not be recognized at all, compared to the input. I now know why, and I guess it just required more knowledge on my side. In the mean time, this is just so; what comes out can't be recognized from what goes in (both graphed). With this Arc Prediction it sure can, because the original samples stay in place. Note though that for OS this is not only about the original samples, but it is even about the "original waves". So, at the rough wave level all has changed. In the end what it comes down to, is that the solution for reconstructing a single frequency, doesn't count at all for complex waves (containing music). Ah, I said that before. Right, but this time my measurement already proves it. What comes out can't be recognized with what goes in. But still it bring you music ?? yes. And thus again : so much can be improved.
Long story short on the last question : I guess it needs quite some knowledge to "just in between the lines improve ripping". But I guess I'm almost ready for it.
One last thing for now and here, and related to the last sentence :
There is more going on than what I currently can reason out 100%;
This is related to hires material not sounding as good as this upsampled 16/44.1; I won't say that this is caused by all hires being upsampled (by the same wrong means) from 16/44.1, but
something is going on with "native" e.g. 24/192 material which makes it hide those on stage sounds and everything. Personally I don't think we can end up at "but Arc Prediction is wrong afterall", just because it does so right for such a long time (5 weeks in my case). Notice the difference with "faked high detail", which is very easy to create by just removing some (lower) mid frequency output. Result : "Ah ! you can hear her spitting in the microphone now !!". Yeah, right. But this won't last for long, because it is "created" by removing something else first. It won't last for 5 weeks.
So, it intrigues me that it is just sheer high resolution I perceive from this Arc Upsampling, while native 24/192 (or 24/96) already shows explicit lower resolution. So, "as good" would already be strange (but according to my own applied theories it can be done), but how "worse" falls into place ? ...
Peter
