@RRod
Chance of long runs of success:
As I said a few pages back I'm well familiar with binomial distribution and I know that with increasing number of trials the chance of seeing any given run length increases asymptotically towards 100%. But my total runs on test192 are maybe high double digits, and I gravitate towards 10% hitting 5% and 1% in places. And the better success (in both tracks) corresponds to tracks where I heard a specific note. Remember there are other tracks where I couldn't find anything to latch onto and I did much worse. To say it could still be chance is always true - anything can always be chance. But at this point I am starting to believe I may be hearing something.
@castleofargh
Lack of previous tests:
I'm glad the error of your ways you have learned, my young padawan. (insert maniacal laugh).
But seriously, THIS is the biggest problem. I (and others) have posted similar questions of various forums of knowledgeable people. Predictably the same thing always happens, the answers quickly sort out into the two camps. The camps agree on absolutely nothing, EXCEPT that they are both so sure of their position they consider testing is a waste of time and any test that is contrary to their position must be flawed. So this is IS a Kobayashi Maru situation for the tester (that was the best reply to my whole thread IMO). It's exactly like the expectation of listening - people that read a test result once a while back remember it's conclusions the way they want. If you go back and actually read the test results, it didn't actually test for that conclusion. I have looked hard and challenged many, and I am coming to believe there is no 24 v 16 test that has proper results posted. That is a problem for the industry.
No dither:
The reason for no dither is simply testing time. I wanted to pass the first one and then make it harder. But the first one was hard enough. I'm going to take a two week hiatus from testing and posting, then I'll be back and probably try dither. Frankly I don't think I stand a chance with dither added, but I need to buck up my courage.
Noise:
I have said repeatedly, but people refuse to believe me, I am not rejecting any of the noise theories. I am well versed in the theory of noise, I know it is around 100 dB etc etc. I'm not willfully ignoring it - I just don't hear it. I don't hear it on CDs I don't hear it on MP3s and certainly not on any of these HD tracks, even with the volume up. I'm using headphones and there is a limit to how loud I am comfortable with.
So what am I hearing?
That's a 64 bit question. I don't know, I think there is scant research on the very edges of audible quantization errors. I don't know why, but I suspect for communications research in ADC they experiment with shaping the noise in different ways to get a better bit error rate, which is relatively easy to measure. Although a lot of the techniques developed for Comms get used in Audio, it's not the same end game. For audio my impression is that most of the research dollars in the 90s was in the psychoacoustic models for compression. To test that, you put 100 average subjects with average or typical consumer listening situations and fiddle the model until they are can't discern. It's not exactly about finding the boundary in the best uncompressed conditions. For people making high end equipment, they all need their secret voodoo sauce or marketing and they have scant interest in funding tests that may prove there is no voodoo and no sauce.
So still what am I hearing?
My best guess lies in a misconception that many have about quantization noise. Please open any standard EE textbook on signals and systems. The first thing you will read is that Quantization errors are a non-linear process and cannot be completely analyzed mathematically. The ~6db per bit idea (which is where you get your 100 dB and 60 dB) is an
approximation. I'm not making this up it says it right there in the textbook. I saw a U of Waterloo paper (google) that had a good intro summarizing Q noise, google it. The analogy is similar to FM and AM radio. In EE you learn how to completely analyze AM using Fourier and Laplace. Every detail can be described by nice equations with precise answers. Then the next thing you learn is that FM is a non-linear process that has no equivalent equation. But it sounds better. This revelation is very frustrating to young padawans, but you get over it after a few weeks. A few tricks and approximations and maybe computer simulations are used to analyze FM to enough extent to be able to use it. Same with Q noise - and actually it has some similarities to FM with repeated short spikes throughout the spectrum.
FFT:
You also can't say conclusively you looked at the FFT and didn't see anything. While I agree it's true that you aren't going to miss some 50 dB spike, there are limitations with FFT. Signals move in time, FFT is a slice in time. To convert between domains you need a window like Hann or Blackman and the windows have artifacts. I think everyone that has worked with this stuff hands on knows this.
Possible Theory:
So remember that a frequency shift and phase shift are the same thing (while they're shifting). My suspicion is that human hearing is incredibly attuned to minute frequency differences, which make up what we call "tone". Note how well we pick out peoples voices, or a Stradavarius, or a Gibson. I'm sure you can pick out Mick Jagger or Bono or Bruce Springsteen in the first syllable. No tones are pure, they all have distortion and we can pick out the slight differences in the higher order harmonics. It may be that at some resonant frequencies the quantization introduces just enough frequency shift that you can detect it, or mucks with the relative amplitude of certain harmonics.
Conclusions:
Well as I said I'm taking a 2 week hiatus but I'll be back with my phasers fully energized to take one last dig at this. I only found one poster that did similar tests, and also inspired me to try. His results seem to match mine. He also said it was very very hard to discern. He could only detect on one of his headphones, the other was pure guess. I may be limited in equipment. I've also thought about other possibilities. The first is that the sample tracks weren't recorded in 24 bit, they were upsampled from 16. That would account for a fail, but not a pass, unless I'm hearing artifacts of the upsampling. Since I am well familiar with upsampling algorithms, I think that is even less likely to be audible than 16 bits, but not non-zero. The other possibility is simply a minor bug or flaw in the 16 to 24 upsampling (if they did that) - also very remotely unlikely but probably more likely than the other theories. Or a bug or resonance point in the multi-segment converter in the DAC that is triggered by either the 16 or the 24. Probably more likely than the rest, but till very unlikely to be audible.
I would very much like to conclude that 24 bit is hooey and move on. However four big things give me pause: (a) lack of solid testing data published on this problem, and other anecdotes similar to mine (b) Engineers and architects at TI, Cirrus/ESS, and Wolfson are not idiots. I've met some of these people they are very serious ridiculously smart and educated people. Why would they all invest so heavily in 24 bit architectures for the last N years and push performance another 30 dB past 100, if there was absolutely no use for it whatsoever. It could be Marketing, I know, but I still pause. I've read their papers they seem to be trying seriously to make the DAC and ADC better. And they are amazing already. (c) People inside Dolby in the 90s, who had no dog in this particular hunt at the time, told me the be all and end all was somewhere in the 18-20 bit range. This was when Dolby AC3 was being commercialized for DVDs. Dolby AC3 on DVD is 16 bit, but you may not know there were 18 and 20 bit AC3 options that were only available on the Pro equipment that Dolby made for the theaters and studios. A friend of mine left Dolby and joined me at another company and we talked about it later and it seems they were doing serious science it wasn't fluff. (d) why would anyone bother with dither if 96 dB was 30 dB too much? This one doesn't seem to be marketing because there is not "dither sticker" on the label of a CD.
I also note that apparently Neil Young posted on the Pono website in the last few days that Warner Brothers has a huge catalog of 24 bit material and he is pushing them to release it. I don't know anything more than that, but if true, 24 bit mania may be upon us soon. It would be nice to know the answer
I take the hiatus with the tentative guesstimate conclusion that the be-all and end-all is somewhere in the 18 bit ENOB area, give or take a bit. Well implemented dither may get you some of the way there, and dither may get you to the point that detection will require such an exotic combination of gear and ears as to render it such that 1 in 1000 people can detect so it probably becomes moot. Certainly a good recording trumps all of this by 100 country miles.