It wasn't for you anyway.And you sound like a broken record….Which is my impression, not a scientific double blind one as I lack reference. I do not know anyone personally who keeps on hammering
How many readers did you poll before getting that conclusion?
A p value of 0.05 means that if you have completely random test results, iow your test data is random, 1 out of 20 of such test results will be "significant". But your results are still random.
So if you do some tests, lets say 10, the chance of getting a "significant" result is, lets calculate that very precisely: pretty effing high.
Repetition of tests is key when finding out the truth. In this case repetition of the test didn't produce a "significant" result.
You guy's are looking only at the one supposedly significant result, I'm looking at all the data.
Edit: Iow one significant test result is not significant.
Last edited:
I would agree with you except for the last two sentences if the retest results were all over the place, like the original test results of the other participants, but in his normal state of consciousness and for the large echoes, Tfive got the exact same results for the retest as for the original test: correct in all cases except for the Saint-Saens recording that he had complained about. That's why I wonder if the test wasn't simply too insensitive - which is only speculation, of course.
Then again, if four pieces of music is "torture", one could hardly expect anyone to listen very carefully to even more recordings. I also think I should have insisted on not using the Saint-Saens recording for the retest, then we wouldn't have had this discussion in the first place. Oh well.
Then again, if four pieces of music is "torture", one could hardly expect anyone to listen very carefully to even more recordings. I also think I should have insisted on not using the Saint-Saens recording for the retest, then we wouldn't have had this discussion in the first place. Oh well.
As always, more testing is necessary.
Or maybe not, 99% of all commercial converter designs are very very likely to be "blameless". There are lots of test results pointing to this, just take a look at the scientific magazines in this subject.
And adding artefacts to a digital system, by for example using reconstruction filters that can do this (for example NOS converters), will eventually lead to audible differences.
Is that important or useful in any way, I don't think so.
Or maybe not, 99% of all commercial converter designs are very very likely to be "blameless". There are lots of test results pointing to this, just take a look at the scientific magazines in this subject.
And adding artefacts to a digital system, by for example using reconstruction filters that can do this (for example NOS converters), will eventually lead to audible differences.
Is that important or useful in any way, I don't think so.
Fair enough, but if you are convinced that commercial DACs are good enough, what did you mean by "amatures like us can make an improvement" in post #1519?
It's a different issue than pre- and post-echoes, but I wouldn't call typical oversampling DACs blameless, as almost all of them clip on peak sample normalized music recordings. Of course music recordings shouldn't be peak sample normalized, but they often are.
It's a different issue than pre- and post-echoes, but I wouldn't call typical oversampling DACs blameless, as almost all of them clip on peak sample normalized music recordings. Of course music recordings shouldn't be peak sample normalized, but they often are.
As so often is the case, the truth probably lies somewhere inbetween...
Bill is correct insofar as modern DAC chips that employ upsampling with an ASRC usually use filters with far less passband ripple. For the ESS9038q2m for example with it's best filter (regarding passband ripple - only 0.002dB) has echoes another 40dB down from the "small echo" files in the test (if I remember my calculations correctly which are somewhere here in the thread). I'm pretty sure I would never be able to hear this at all.
For me it's far more likely that NOS DACs produce a significantly different sound signature as sampling rates are changed. OS DACs will probably largely circumvent this phenomenon by running their DAC output stages at a constant sampling rate by using a form of ASRC. That's still my theory at least.
Bill is correct insofar as modern DAC chips that employ upsampling with an ASRC usually use filters with far less passband ripple. For the ESS9038q2m for example with it's best filter (regarding passband ripple - only 0.002dB) has echoes another 40dB down from the "small echo" files in the test (if I remember my calculations correctly which are somewhere here in the thread). I'm pretty sure I would never be able to hear this at all.
For me it's far more likely that NOS DACs produce a significantly different sound signature as sampling rates are changed. OS DACs will probably largely circumvent this phenomenon by running their DAC output stages at a constant sampling rate by using a form of ASRC. That's still my theory at least.
ESS uses an ASRC by default, but other OS dac manufacturers don't do that. Most of them, including ESS when in synchronous mode, run the OS output at some multiple of the input sample rate by using an internal MCLK divider to produce the required output clock frequency.
OK, then it's not an ASRC anymore but only a SRC. AFAIK technically not much different, it also has to involve digital filtering much like the ASRC.
It can take a lot of repeated listening to notice some new aberration music recordings. If one listens to a lot of music with buried echoes, the brain may learn to ignore them, or it may learn to pick them out as objectionable (assuming in either case the echoes are above some combination of reproduction system and listener thresholds).
A more proper way to conduct testing would be to train people to hear low level echoes by starting with easily heard high level echoes. Then reduce the echo level and give people immediate feedback when they successfully hear or else fail to hear echoes. Over time they should learn to detect smaller and smaller echoes.
When their abilities stop improving would be the time to start blind testing them.
Otherwise there may be some people who are already very sensitive to echoes, other people who have subconsciously learned to ignore them, etc. This especially can be a problem when there are a small number of test subjects.
Regarding listening torture, that term describes what I had to do to sort recordings of audio opamps in order of distortion by ear. It was very hard to do. Most people probably wouldn't to do that to themselves. Not sure I would want to do it again either. However, if anyone likes that kind of torture, I saved the opamp test files. Happy to make them available if anyone wants some of that. ...What, no volunteers?
A more proper way to conduct testing would be to train people to hear low level echoes by starting with easily heard high level echoes. Then reduce the echo level and give people immediate feedback when they successfully hear or else fail to hear echoes. Over time they should learn to detect smaller and smaller echoes.
When their abilities stop improving would be the time to start blind testing them.
Otherwise there may be some people who are already very sensitive to echoes, other people who have subconsciously learned to ignore them, etc. This especially can be a problem when there are a small number of test subjects.
Regarding listening torture, that term describes what I had to do to sort recordings of audio opamps in order of distortion by ear. It was very hard to do. Most people probably wouldn't to do that to themselves. Not sure I would want to do it again either. However, if anyone likes that kind of torture, I saved the opamp test files. Happy to make them available if anyone wants some of that. ...What, no volunteers?
Last edited:
To be honest, in this case ignorance is bliss.
As I stated, if I can't hear echoes of -40dB reliably why should I train myself to hear echoes at -80dB? I'm in for the music und not for the spotting of the minutest differences.
In this regard I really don't envy you for your trained ears. Must be a torture for you to listen to almost any system out there.
Me on the other hand , I regard myself more as a musician and a music lover. I have trained ears in a very different sense, can spot minor musical details and can admire musicians for producing them, even on very mediocre systems. And I learned a lot when mixing/mastering my band's albums about tonal and volume balance.
I already told Marcel this and I will repeat it here: I would rather listen to any of the files with echo instead of not being able listening to them at all (provided the music hits my personal taste, which is a whole other story).
Or as Marcel put it - he would prefer a wax cylinder of Caruso any time over a highres recording of Britney Spears 😀
As I stated, if I can't hear echoes of -40dB reliably why should I train myself to hear echoes at -80dB? I'm in for the music und not for the spotting of the minutest differences.
In this regard I really don't envy you for your trained ears. Must be a torture for you to listen to almost any system out there.
Me on the other hand , I regard myself more as a musician and a music lover. I have trained ears in a very different sense, can spot minor musical details and can admire musicians for producing them, even on very mediocre systems. And I learned a lot when mixing/mastering my band's albums about tonal and volume balance.
I already told Marcel this and I will repeat it here: I would rather listen to any of the files with echo instead of not being able listening to them at all (provided the music hits my personal taste, which is a whole other story).
Or as Marcel put it - he would prefer a wax cylinder of Caruso any time over a highres recording of Britney Spears 😀
Which is, unfortunately, something that we do not have the luxury of. It's disappointing that the thread is approaching 80K views, yet we struggle to persuade 6 participants to voluntarily conduct the relatively easy to perform 176.4 upsampling experiment.
Bill,
Notions of 99% of converters being blameless ignores the fact that most commercial OS converters utilize an objectively compromised on-chip interpolation-filter. Which are typically quite far from the ideal SINC-function based interpolation of the sampling theorem. Compromised, in terms of it approaching the band-limiting ideal SINC function, without introducing implementation artifacts. One of our earlier experiments indicated (with an admittedly too low level of statistical confidence due to the participation numbers problem mentioned above), that bypassing typical on-chip OS filters, and instead utilizing an high-performance external OS filter essentially produces the same subjective sonics as simply omitting all OS interpolation filter altogether (NOS).
Quite interestingly, it seems that the nearer an OS filter approaches the ideal SINC-function, the more it sounds like NOS, and the less it sounds like common on-chip OS. Which, of course, points the finger of guilt at common on-chip OS filters. Such filters appear to be introducing some sort of audible artifact, making them sound different from either NOS, or from external high-performance OS.
Marcel's echo test was for the purpose of attempting to isolate and identify one particular potential causal mechanism. There are others. There's no question that equiripple FIR filter time-domain signal echoe/reflection artifacts exist, as Lagadec and Stockam showed long ago. The question is whether such echoes are what's causing the characteristic subjective difference between OS and NOS. All of the above, of course, presumes, that a participant hears a characteristic difference between NOS playback, and common on-chip OS playback.
"...have trained ears in a very different sense, can spot minor musical details and can admire musicians for producing them, even on very mediocre systems. And I learned a lot when mixing/mastering my band's albums about tonal and volume balance."
Been there, have done all that too. Did it for a living for 7 years, right after of college.
Been there, have done all that too. Did it for a living for 7 years, right after of college.
Last edited:
Anyone can be a scientist, if you use the scientific method.
Unfortunately, not everyone who calls him/herself a scientist is an actual scientist.
Then your using your equipment outside its design range, audible artefacts are expected. Inside their operating range, they are blameless.
Its very easy to design a dac that doesn't clip audible with normal music, use a bit of headroom for the analog part of the dac.
That is indeed unfortunate.
The reason I don't participate in these tests is that I'm 51 and my ears just don't work as they did 40 years ago. I's a young peoples game to spot tiny differences.
If I look at the noise level, spectrum of the noise and harmonic distortion numbers of converters, even the worst have thd+noise at -80dB fsl.
And my experience tells me that it's really difficult to hear artefacts that low below the signal level with normal signal strength. Masking is a very real thing. And although our hearing system can detect sounds with a change in spl of 120dB. Our hearing system can't do that listening to a sound at 120dB spl and then one second after that hear a sound at 50dB spl, it just can't happen. We actually have to adapt to very soft sounds. Just see a youtube video of people in a silent anechoic room. It takes 10-15 min to hear your blood flow through your ears. Lower than that we can't hear, that's the theoretical limit. Again masking is very real.
But if anyone shows me some evidence of what you just claimed, i'll shout it from the rooftops. So far I haven't seen this in the scientific literature.
Last edited:
I sure as hell won't call myself a scientist, I'm just a stupid engineer.
However, what I was curious about is what there is to improve about DACs in your opinion.
The analogue part is usually not the problem, it's the digital filters that are very often designed with zero headroom. Of course you can make your own filter with an FPGA to avoid the issue, which is precisely what I have done.
That’s a big misconception.
The older I became the better and faster I could tell one system from another.
I can hear all kind of differences despite not having the ears of an 18 year old, so loss of HF is not as bad as you might think. Your brains are a very important part in what you hear.
Youngsters are generally not interested in this sort of things. To them sound quality is mostly no issue at all, volume is way more important.
So I see no reason for you not to participate.
When noise level is -80 dB, it doesn’t mean that hearing sound stops at this level.
Like an FFT, our hearing system has filter bins that are 3 to 4 Hz wide in the sensitive areas.
So when noise over a 10Khz BW has a level of -80dB, the level at the 3Khz bin is 10log(4Hz/10Khz) = circa 30 dB lower.
That’s why we can hear a 3khz tone at -100dB that is not masked in -80dB broadband noise.
Hans
Last edited:
I suspect that you may be missing the point of what we are investigating, which helps explain why we are doing it this way. I think the first thing to recognize is, a participant either hears an characteristic difference between OS and NOS, or they don’t. If your 51 year old ears don’t hear this difference, then you are correct about not participating in our tests. By the way, my ears will be 60 years old next month, and they do hear the difference, because the effect seems to apply across the entire audible band, including the bass range. May I suggest, that you perceive with you ears, as well as read with your eyes, in determining what constitutes evidence you should consider.
For those of us who do hear a characteristic OS/NOS difference in playback, datasheet specifications don’t seem to readily explain why. Not that they don’t necessarily contain that objective information, but because we don’t know, as yet, which specifications will positively indicate the OS/NOS difference in sound character which we hear. (Aside from the ZOH based treble droop of NOS, which is easily analog EQ’d). Prime to identifying what my be the relevant datasheet specifications, is to first determine which functional block is responsible. It appears we have determined that it is the typical on-chip OS interpolation-filter design of most DACs which seems somehow responsible.
I'm pretty sure you are not, but the way you are able to organize results and observations makes you look like a pretty damn good one. I've seen work from recognized scientists that look like a pile of junk that only they could eventually understand (sometimes).
An other umpa lumpa of science 😀 (one of Sheldon Coopers best jokes)
I'm interested in good sound.
I look at what inhibits this for the most. I think this is speakers by a very large margin, so I'm focussing on that.
An analogy: If you want to loose weight, you can shave of your pubic hear and you will indeed loose weight. But not much (I hope). It's a better strategy to eat one sandwich less each day.
Modern commercial DAC's of a couple of hundred bucks perform at the theoretical limit of what is possible. Artefacts are 120dB below the full scale signal level. I just can't see how anyone can improve upon that, without serious cooling equipment.
And then comes the question of, is that really necessary for good sound?
I don't think so. Vinyl can give you really good sound (if you disregard the clicks) and has artefacts 60dB higher than a decent dac. That is not a lot more, that is a whole lot more.