The No.1 Website for Pro Audio
88.2 - 96kHz Vs 44.1-48kHz (a thread to end them all!!)
Old 22nd January 2020
  #691
Listen first

Tom Barnaby, it is time for you to stand down.

For the record, I am in agreement with you that 96k recording has practical value for some kinds of program material. I am on record recommending that more engineers adopt 2x rate recording as standard practice. Now please stop.

Your arguments in favor of our shared position are not helpful. They are full of fundamental misconceptions about basic sampling theory. They are confusing to people who don't know that theory intimately and maddening to those of us who do. Fourier analysis works, and Nyquist was not wrong. These theorems are as true today as the day they were penned, but both are capable of being misapplied by people who don't understand their postulates and implications. The fact that audio reproduction is still imperfect is no reason to blame the mathematicians whose work underlies so much of modern electrical engineering. Likewise, if one day my house burns down, I won't conclude that Ohms Law is wrong.

Quote:
Originally Posted by Tom Barnaby View Post
What did I say that is full of nonsense?
I read through everything you wrote in this thread. It's pretty much all wrong. You clearly don't know enough math to be arguing these points at all.

Quote:
Please, explain your point of view as clearly as you can .
I am ready to learn.
I can't possibly teach a college class on linear systems theory in a forum post, and nobody wants to read it either. But that's the minimum background required to truly understand how the work of Fourier, Nyquist, and Shannon applies (or doesn't) to digital audio. So when the people on this thread who do have such background (Hi Fabien!) try to explain the theory to those who don't, they're forced to explain it in words rather than in mathematical notation. Please understand that what you're getting then is an imprecise analogy to the actual mathematics. If you find fault with it, that doesn't mean the underlying math is wrong.

We soldier on because some people here really do want to understand the basic theory as best they can, rather than inventing half-baked explanations of their own. I've even done my best to explain some "non-Nyquist" sampling theory in earlier posts on Gearslutz -- do a search if you're interested.

I'll close by saying this: If you want to understand why I think high-rate digital audio has practical value, my answer is "first hand experience". If you'd like to know why I think my ears hear what they hear, it seems to me two facts suffice:
  1. Real-world sounds are not brick-wall band-limited
  2. Human hearing is not perfectly linear.

There's no need to drag Nyquist et al. through the mud because of either point.

David L. Rick
Old 22nd January 2020
  #692
Gear Guru
 

Quote:
Originally Posted by David Rick View Post
Tom Barnaby, it is time for you to stand down.

For the record, I am in agreement with you that 96k recording has practical value for some kinds of program material. I am on record recommending that more engineers adopt 2x rate recording as standard practice. Now please stop.

Your arguments in favor of our shared position are not helpful. ...
I read through everything you wrote in this thread. It's pretty much all wrong. You clearly don't know enough math to be arguing these points at all.

We soldier on because some people here really do want to understand the basic theory as best they can, rather than inventing half-baked explanations of their own....
David L. Rick
Big Ups to Rick, Fabien and rest of the humans who DO understand digital audio for continuing to patiently put out the Brushfires of Bull$#!t that keep popping up on these topics.

Sadly, I fear you are never going to convince the Dunning-Krugers, the trolls, the Audio Flat-Earthers, etc. They are write-offs.

But the open-minded people who are browsing and lurking deserve to see the bull$#it challenged whenever it rears its ugly head. Otherwise, ignorance spreads.
Old 22nd January 2020
  #693
Gear Maniac
 

Quote:
Originally Posted by David Rick View Post
Tom Barnaby, it is time for you to stand down.

For the record, I am in agreement with you that 96k recording has practical value for some kinds of program material. I am on record recommending that more engineers adopt 2x rate recording as standard practice. Now please stop.

Your arguments in favor of our shared position are not helpful. They are full of fundamental misconceptions about basic sampling theory. They are confusing to people who don't know that theory intimately and maddening to those of us who do. Fourier analysis works, and Nyquist was not wrong. These theorems are as true today as the day they were penned, but both are capable of being misapplied by people who don't understand their postulates and implications. The fact that audio reproduction is still imperfect is no reason to blame the mathematicians whose work underlies so much of modern electrical engineering. Likewise, if one day my house burns down, I won't conclude that Ohms Law is wrong.


We soldier on because some people here really do want to understand the basic theory as best they can, rather than inventing half-baked explanations of their own. I've even done my best to explain some "non-Nyquist" sampling theory in earlier posts on Gearslutz -- do a search if you're interested.

I'll close by saying this: If you want to understand why I think high-rate digital audio has practical value, my answer is "first hand experience". If you'd like to know why I think my ears hear what they hear, it seems to me two facts suffice:
  1. Real-world sounds are not brick-wall band-limited
  2. Human hearing is not perfectly linear.

There's no need to drag Nyquist et al. through the mud because of either point.

David L. Rick

Hi David.

I don't think that anybody in this thread tried to drag Nyquist, Shannon or any other scientist through the mud.

They are very respected scientist and we all know that their contribution to modern technology was very important.
However, questionning admitted theories is essential in science and everybody should have the right to ask questions.

This subject is difficult to debate because the mathematical background is very complex and most of us are not experts in this area.
However, I keep thinking that listening experience is important and should not be neglected, even if it contradicts globally admitted theory.

I had the opportunity to listen to music recorded at 352.8 kHz and, for me, it doesn't sound like music recorded at 44.1 kHz or even 96 kHz.
Old 22nd January 2020
  #694
Quote:
Originally Posted by studio96 View Post
Now we are back at the very beginning of my first post. This is what people regularly stumble about and tend to overlook. They just mention Nyquist's theorem, are happy but d do not apply it to practical cases:



1) Not all signals we have in audio signal processing or generation are really continuous. Square waves and triangles e.g. are not - especially when used as phase modulators. In theory they are too and can be described that way but practically this description requires and infinite iteration which cannot be fulfilled. So even the theory of separation practically fails. Real hardware additionally comes always with a particular band limitation.

And even a perfect and clean sine wave cannot totally be described or handled this way perfectly, when it starts with full amplitude. Even when it starts at Zero and the first sample is zero too, it's d/dt is not continuous.

2) There is no filter giving this "proper band limiting" you are demanding. Neither digital nor analog filters can do this. There also is no known filter totally having a Zero in the stop band at all, so it is not possible to practically reconstruct everything perfectly. It is always imperfect. And the amount of "imperfectness" mostly has to do with the frequency to be reconstructed.

Or in short words: Theory is right, but you cannot reach it.


1) and 2) are working totally fine. You really just raise totally illogical assumptions ("your idea is nice, but it lacks the rainbow rocket driven unicorn!"):

1) An audio signal generally is continuous. No matter its shape, it is fully decomposable. Any physical event in the universe is band-limited. It isn't a matter of choice. There is no perfectly sharp triangle or square in the whole universe. It's rather unproductive to assume it, much like insisting on the presence of unicorn before proceeding.

2) You don't need an infinite stop band attenuation. Although even these exist in certain situations, called "null poles" or "zeroes". According to the theorems, you can even achieve infinity, given you're willing to deliver infinite resources as well Fair play! No?

A digital signal/system is specified in terms of bandwidth and noise. In this universe, bandwidth can't be zero or infinite, noise can't be infinitely low either. Both the sampling and Fourier theorems take these realities into account. And if you wouldn't trample over these details, you'd understand how clever the theorems are, and why they work so well. This stuff is made for our specific world, not Netherland or Tatooine

Understanding that just any event has some noise and a well defined bandwidth, is the key to Fourier and sampling. No events are free of noise and/or have infinite bandwidth, infinite rise time, infinite energy. This type of universe doesn't exist. Much like rocket unicorns never existed in the first place.

Modern filtering technology can easily attenuate down to the 24bit noise floor, within a fraction of an octave. The AD/DA next to you does it effortlessly. And if you need even more, just give the machine more power.

Pls post an audible signal that can be represented at 88.2kHz, but not with 44.1kHz. You say it's easy? Then do it pls, hosting is free. Because in this moment, a gazillion ADs, DAs and SRCs are working just fine (with music being just a niche subset of many "more serious" "far more sensitive" applications).

Last edited by FabienTDR; 23rd January 2020 at 12:34 AM..
Old 22nd January 2020
  #695
I think this type of skepticism (it's fine, really) has its roots in the analogue sampling/processing heritage, where nothing is really deterministic, typically subject to relatively huge tolerance (enormous in comparison with digital).

Analogue gear is, by definition, "never completely there". Constant skepticism, verification try and error and fine tuning are considered as good manners (for good reason).

But digital is built on a totally different mind of thought, where determinism makes the central pillar. Digital is fully predictable, today and tomorrow, even by a 12 year old coder. And it's this "wonder" that troubles so many. No need for try and error, it's deterministic.

Last edited by FabienTDR; 22nd January 2020 at 10:41 PM..
Old 22nd January 2020
  #696
Gear Addict
 

I have some interposed question. I see it purely from the perspective to reduce aliasing while processing. As I've understood it aliasing is one of the most important things that reduce sound quality for in the box saturation, right?
  1. What what do you recommend nowadays as a sampling rate in the DAW. I''m using a lot of plugins also for saturation and since I've had a look in the analyser I see how many of them create aliasing.
  2. Some of them do oversample to reduce it, but is SR really transparent for the human ear? Or is this really a solution for the problem and you can run your project at 44/48khz without any disadvantage?
  3. What do you think about using plugin shells that can run VSTs with 88/96khz inside a 44 khz project to have oversampling for plugins that doesn't support it?
  4. Does using 48 khz already have real world advantages to reduce aliasing in plugins or is it too near to 44 khz to help noticeably?
Old 22nd January 2020
  #697
Quote:
Originally Posted by DeadPoet View Post
I'm trying to learn about the subject. Would you care to share your thought on the "plugins that don't upsample alias at 44/48kHz" vs "speakers create IMD at 88/96 and up" dilemma? Is there one that is a lesser evil in your opinion?
I wouldn't say that "plugins that don't upsample alias at 44/48kHz", that's simply not true. But plugins that apply fullband nonlinearities at a rate of 44.1kHz or 48kHz (e.g. wideband saturators or dynamics) will likely produce audible aliases.

Aliasing is one form of IMD. In that sense, both can produce very similar audible effects. Aliasing can be seen as an intermodulation with Nyquist. Both sound rather hard, grainy, rough and harsh, both break harmony.

You minimize both by using processors that clean up their own "mess". Any plugin that bandlimits whatever it adds will also automatically remove ultrasonic content. Making sure it can't run into another nonlinearity and produce more IMD. That's optimal.

What isn't great is when plugins don't respect the most fundamental condition of the sampling theorem (bandlimiting). Only in this case you start to see aliasing, you fight it with manual oversampling, with ultrasonic content in turn potentially provoking more IMD as it hits the next nonlinearity (which wouldn't have happened using an economical rate, and properly bandlimited processors).
Old 23rd January 2020
  #698
Gear Guru
 

Quote:
Originally Posted by FabienTDR View Post
I think this type of skepticism (it's fine, really) has its roots in the analogue sampling/processing heritage, where nothing is really deterministic, typically subject to relatively huge tolerance (enormous in comparison with digital).

Analogue gear is, by definition, "never completely there". Constant skepticism, verification try and error and fine tuning are considered as good manners (for good reason)..
I can remember that skepticism very clearly. I would even go so far as to call it paranoia. Every playback of every take was guilty until proven innocent and if you did not approach it that way, you were not really doing your job. No matter how much of a hurry you were in, you played it back. God forbid you let something slip past and did not detect it until after the musician went home. How many times I heard something a little 'fishy' on a guitar overdub, cleaned the heads, found a speck on the Q-tip - and what do you know, the next pass was perfect.

IMO, it's actually a very fine line between "leaning in" to hear the tiniest of differences and just flat out imagining those differences. Nobody is "immune" to bias. I believe the best engineers are the ones at least understand how fine that line is.

Quote:
But digital is built on a totally different mind of thought, where determinism makes the central pillar. Digital is fully predictable, today and tomorrow, even by a 12 year old coder. And it's this "wonder" that troubles so many. No need for try and error, it's deterministic
Indeed, some people seem to poke at their DAWs as if they were black boxes built by Aliens. Black boxes that no human understands how they work.

What's oddest to me is that the people who most exhibit this paranoia about digital tend to be the ones who started recording with digital. I know very few old-time Tape Guys who go around believing they can hear the difference between two files that null for example. Or who think that uploading a file to the internet "changes the sound". I can remember threads with people swearing that if you consolidated an edited track , it sounded "worse"!

It's as if people need to have some Gremlins to worry about, and if there aren't any, they will make some up.
Old 23rd January 2020
  #699
"black boxes that no human understands how they work"

Is a great point imho. The whole idea to digitally sample a continuous signal is somewhat counter intuitive. And established explanations often confuse with simplifications.

One popular misconception is the idea that digital sampling is taking regular, infinitely small probes of the signal. In theory, this could work under strict "lab" conditions. Sadly in reality, this is a completely misguided way to think about it.

As a consequence of logic, but without being explicitly mentioned by the theorem, these probes aren't exactly of a small size. In fact, each probes is optimally infinitely large, both into past and future. To visualize the point, following the band-limiting requirements of the sampling theorem, this is the exact impulse response, i.e. the shape/weighting of each probe the theorem insists on:



This is the sinc function, basically the atom of digital sampling (the "wobbles" extend to infinity both left and right).

When your AD records audio, it doesn't just probe the immediate level, it also takes into account all past and future values, weighted by the sinc "curve". It quickly diminishes over time, but each probe is really taken over an infinitely wide period!

The same happens on playback. A single sample in your PCM signal, say, a "1", translates to something like this:



On playback, the final reconstruction will be a weighted sum of past, momentary and future samples (in the PCM data). Not just the momentary PCM value.

This sinc curve (technically, imposed by the Nyquist filter) is the glue. This is the magic conversion method from "digital" to continuous, and the vice versa.



Disclaimer: I mention a lot of weird things above, involving "all future samples". Please don't try to use this as a lever for criticism. There are very practical, very effective solutions to all of these, but not mentioned to prevent complication. I'm just trying to help get digital sampling into perspective.

Last edited by FabienTDR; 26th January 2020 at 11:09 AM..
Old 23rd January 2020
  #700
Lives for gear
 
doom64's Avatar
While you all were still arguing about 44.1 and 192 I was running tracks through nice analog gear... courtesy of Access Analog and mix:analog.

Gear that rounds off transients, adds harmonics and (mostly) transparently squishes signals so that they are upfront and alive. Sounds really good (except the raised noise floor) and was recorded with modest equipment. Digitally recorded? Yes...it is my preferred medium for many reasons.

24/96 for those wondering.
Old 23rd January 2020
  #701
Lives for gear
 
chrischoir's Avatar
 

Quote:
Originally Posted by AreYouHuman View Post
What what do you recommend nowadays as a sampling rate in the DAW
60 is optimal but since converters are 88 or 96 then those are recommended
Old 23rd January 2020
  #702
Gear Guru
 

Quote:
Originally Posted by FabienTDR View Post
"black boxes that no human understands how they work"

Is a great point imho. The whole idea to digitally sample a continuous signal is somewhat counter intuitive.

Yes! I do think the counter-intuitive part is the crux of the difficulty people have. The idea that you can define something as apparently "busy" as a sine wave using only two points does not conform to our everyday experiences.

Quote:
And established explanations often confuse with simplifications.
Which is probably why so many people believe in the "stairsteps"
Old 24th January 2020
  #703
Gear Addict
 

Quote:
Originally Posted by chrischoir View Post
60 is optimal but since converters are 88 or 96 then those are recommended
is 60 a typo or is there something I don't get right?


Otherwise I'm wondering a bit why my question above seems not to be interesting. Are they 'stupid' or if they have been answered often before and I've overlooked it? Or just not interesting for others? :-)
Old 24th January 2020
  #704
Lives for gear
 

Quote:
Originally Posted by AreYouHuman View Post
is 60 a typo or is there something I don't get right?
60khz was suggested by dan lavry in his famous white paper - dunno from what year this dates and if other engineers brought this up before him.

Last edited by deedeeyeah; 25th January 2020 at 08:32 AM.. Reason: typo
Old 24th January 2020
  #705
Gear Addict
 

Thanks for the clarification!
Old 25th January 2020
  #706
Lives for gear
 
chrischoir's Avatar
 

Quote:
Originally Posted by AreYouHuman View Post
is 60 a typo or is there something I don't get right?
yes, 60khz is preferred . But as stated previously it is not available in any DAW or with any converter, so you use 88 or 96

Quote:
Originally Posted by AreYouHuman View Post
Otherwise I'm wondering a bit why my question above seems not to be interesting. Are they 'stupid' or if they have been answered often before and I've overlooked it? Or just not interesting for others? :-)
I don't think they are stupid questions at all, but on GS your questions will cause lots of controversy between/among all the experts. Sampling is a touchy subject. You have 3 camps/. One camp of people that are very literal, one group that are more musical/artistic and then a third who are more practical. There is always some overlap and it causes great confusion.
Old 25th January 2020
  #707
60 kHz sample rate (and why we shouldn't bother)

Quote:
Originally Posted by deedeeyeah View Post
60khz was suggest by dan lavry in his famous white paper - dunno from what year this dates and if other engineers brought this up before him.
Lavry's suggestion of 60 kHz was based on avoiding the significant passband droop which would result from first-order analog filters at 20 kHz [Lavry,2012]. This is not a particularly convincing argument: nobody designs first-order analog anti-aliasing filters. Lavry knows this of course. The main argument in his paper is that 192k sampling is unnecessary and leads to reduced converter linearity. He does not argue that 44.1 kHz is sufficient and seems to accept 96k as a reasonable compromise for compatibility with existing standards. His only complaint is that it's a slight waste of channel capacity.

A similar suggestion appears in an AES paper published in 1996 by Peter Craven and Michael Gerzon. Their preferred choice is a sample rate of about 66 kHz. Why? Because they assumed a 10 kHz filter transition bandwidth and Michael could hear out to 23 kHz! Eliminate that interesting quirk, and we're left with the same suggestion that Dan Lavry made 16 years later.

Like Lavry, Craven and Gerzon conclude that 96 kHz will be used in practice for compatibility with existing standards. But instead of clutching their pearls about the wasted bits, they exhibit a lossless coding scheme based on third-order linear prediction that reduces the bit rate penalty to a mere 15%. [Craven & Gerzon, Lossless Coding for Audio Discs, JAES 1996]

Sadly, Michael Gerzon died that very same year, but Peter Craven continued to develop these ideas. MLP compression and the MQA streaming format are both outgrowths of this work.

David L. Rick
Old 25th January 2020
  #708
Lives for gear
 
chrischoir's Avatar
 

Quote:
Originally Posted by David Rick View Post
Lavry's suggestion of 60 kHz was based on avoiding the significant passband droop which would result from first-order analog filters at 20 kHz [Lavry,2012]. This is not a particularly convincing argument; nobody designs first-order analog anti-aliasing filters. Lavry knows this of course. The main argument in his paper is that 192k sampling is unnecessary and leads to reduced converter linearity. He does not argue that 44.1 kHz is sufficient and seems to accept 96k as a reasonable compromise for compatibility with existing standards. His only complaint is that it's a slight waste of channel capacity.

A similar suggestion appears in an AES paper published in 1996 by Peter Craven and Michael Gerzon. Their preferred choice is a sample rate of about 66 kHz. Why? Because they assumed a 10 kHz filter transition bandwidth and Michael could hear out to 23 kHz! Eliminate that interesting quirk, and we're left with the same suggestion that Dan Lavry made 16 years later.

Like Lavry, Craven and Gerzon conclude that 96 kHz will be used in practice for compatibility with existing standards. But instead of clutching their pearls about the wasted bits, they exhibit a lossless coding scheme based on third-order linear prediction that reduces the bit rate penalty to a mere 15%. [Craven & Gerzon, Lossless Coding for Audio Discs, JAES 1996]

Sadly, Michael Gerzon died that very same year, but Peter Craven continued to develop these ideas. MLP compression and the MQA streaming format are both outgrowths of this work.

David L. Rick
great post Dave thanks for the clarification
Old 25th January 2020
  #709
Quote:
Originally Posted by AreYouHuman View Post
I see it purely from the perspective to reduce aliasing while processing. As I've understood it aliasing is one of the most important things that reduce sound quality for in the box saturation, right?
Aliased components are inharmonic, thus quite objectionable to most ears. But tiny amounts of aliasing are sometimes used as a brightening effect.

Quote:
What what do you recommend nowadays as a sampling rate in the DAW. I''m using a lot of plugins also for saturation and since I've had a look in the analyser I see how many of them create aliasing.
I'm on record recommending 96k for most audio production. But I don't do a lot of production in which I really crush everything. You might want to put a low-pass filter at the front of your plug-in chain. I think Fabien used to offer a free one.

Quote:
Some of them do oversample to reduce it, but is SR really transparent for the human ear? Or is this really a solution for the problem and you can run your project at 44/48khz without any disadvantage?
I find audio quality to be compromised at 48k and especially at 44.1k, but I'm recording superb musicians playing real instruments in good halls and have high expectations for playback fidelity. Are you seriously concerned about the "transparency" of a compressor in "nuke" mode?

Quote:
What do you think about using plugin shells that can run VSTs with 88/96khz inside a 44 khz project to have oversampling for plugins that doesn't support it?
Do this on a few channels and soon it's more costly (in CPU cycles) than just running the whole project at 96k.

Quote:
Does using 48 khz already have real world advantages to reduce aliasing in plugins or is it too near to 44 khz to help noticeably?
It won't help with aliasing induced by intentionally non-linear processing. You need to be at 2x sample rate to get an improvement. But I do like 48k much better than 44.1k for acoustic stuff that isn't being squashed.

David L. Rick
Seventh String Recording
Old 25th January 2020
  #710
Lives for gear
Quote:
Originally Posted by David Rick View Post
You might want to put a low-pass filter at the front of your plug-in chain. I think Fabien used to offer a free one.
Would that be the TDR Ultrasonic Filter (alpha version)? I'm currently using the also free Brainworx bx_cleansweep V2, but I haven't actually made a shoot-out with hp and lp filters. Maybe it's time I do.
Old 25th January 2020
  #711
Lives for gear
Quote:
Originally Posted by studio96 View Post

2) There is no filter giving this "proper band limiting" you are demanding. Neither digital nor analog filters can do this. There also is no known filter totally having a Zero in the stop band at all, so it is not possible to practically reconstruct everything perfectly. It is always imperfect. And the amount of "imperfectness" mostly has to do with the frequency to be reconstructed.

Or in short words: Theory is right, but you cannot reach it.
While that may be true, from a practical standpoint this doesn't change much. If invariable signals that undergo a null testing process (So one gets it polarity switched, so it mirrors the other signal) in the analog domain, do indeed null to the extent the remnants (the imperfections) need massive gain to even show up on an Analyzer, then these imperfections are of very little interest.

Last edited by Mikael B; 25th January 2020 at 02:22 PM..
Old 25th January 2020
  #712
Lives for gear
Quote:
Originally Posted by Tom Barnaby View Post
We should probably do an analogue nulling test on a spectrum going from 20 Hz to 192 kHz.
How could we do that ?
My suggestion is to do it in the analog domain. Analog equipment actually determines the possible bandwidth, doesn't it? It's likely uninteresting to test whether 44.1 kHz reproduces frequencies above 22 kHz well as few people would expect that, so most quality equipment with a low S/N ratio would probably do. Obviously, if you want to test above the hearing range you must use equipment that at least has very similar imperfections for material above 22 kHz. That's likely another ball game. I'd only be interested to compare higher sample rates with 44.1.

The analogue equipment would likely need to be calibrated separately, at least if nulls will not occur with signals from DAC. As would the digital test material.

As long as you can
  • Use 2 audio interfaces of the exact same model at the same time
  • these can be calibrated so audio is in phase lock
  • sum these in the analogue domain without unacceptable side-effects

then you might have a useful test. I'd assume that calibration would mean testing with waveforms with no overtones to get the time domain right with no phasing. This might be hard — I actually don't know — but is probably not impossible. Once phase can be determined to not be a problem at 44.1 and the other at another value, I'd assume you could view the testing equipment setup as "calibrated".

Then you'd do longer runs of playing different test material — naturally such that do null in the digital domain when one is converted to the other's sample rate — running though this. I'd probably use a sensitive high-quality Oscilloscope filmed as the record of the tests. I'd assume such an Oscilloscope or frequency analyser can be rented. But as any deviation from silence is a result (above equipment noise) a sensitive meter might be informative enough at a first run.

At least this is how I imagine a starting point of a useful test. Likely, during setting this up one might hit snags that need solutions and rethinking. I'm sure greater and more experienced minds than mine see a few issues already.
Old 25th January 2020
  #713
Lives for gear
Quote:
Originally Posted by joeq View Post
I know very few old-time Tape Guys who go around believing they can hear the difference between two files that null for example. Or who think that uploading a file to the internet "changes the sound".
Well, to be fair uploading a piece of music does change the sound, partly due to lossy re-encoding, which is common, as well as the fact that the playback system is changed to one beyond our control. Of course, the lossy re-encoding is not absent from most digital publishing environments — even though there are lossless streaming and outlets — and the playback systems will almost always be beyond what producers can control, so it's not that different from any other publishing.
Old 25th January 2020
  #714
Gear Maniac
 

Quote:
Originally Posted by Mikael B View Post
My suggestion is to do it in the analog domain. Analog equipment actually determines the possible bandwidth, doesn't it? It's likely uninteresting to test whether 44.1 kHz reproduces frequencies above 22 kHz well as few people would expect that, so most quality equipment with a low S/N ratio would probably do. Obviously, if you want to test above the hearing range you must use equipment that at least has very similar imperfections for material above 22 kHz. That's likely another ball game. I'd only be interested to compare higher sample rates with 44.1.

The analogue equipment would likely need to be calibrated separately, at least if nulls will not occur with signals from DAC. As would the digital test material.

As long as you can
  • Use 2 audio interfaces of the exact same model at the same time
  • these can be calibrated so audio is in phase lock
  • sum these in the analogue domain without unacceptable side-effects

then you might have a useful test. I'd assume that calibration would mean testing with waveforms with no overtones to get the time domain right with no phasing. This might be hard — I actually don't know — but is probably not impossible. Once phase can be determined to not be a problem at 44.1 and the other at another value, I'd assume you could view the testing equipment setup as "calibrated".

Then you'd do longer runs of playing different test material — naturally such that do null in the digital domain when one is converted to the other's sample rate — running though this. I'd probably use a sensitive high-quality Oscilloscope filmed as the record of the tests. I'd assume such an Oscilloscope or frequency analyser can be rented. But as any deviation from silence is a result (above equipment noise) a sensitive meter might be informative enough at a first run.

At least this is how I imagine a starting point of a useful test. Likely, during setting this up one might hit snags that need solutions and rethinking. I'm sure greater and more experienced minds than mine see a few issues already.


This should be done comparing two digital recordings, one at 48 kHz and the other at 96 kHz. It should involve actual instruments recordings made in a physical room like a piano or a voice.
Comparing waveforms froms a synthesizer would not be relevant as using 96 kHz is mainly about capturing reverberated signals in a more accurate manner.

Variations of the signal in phase would be limited to the sampling period, so we would need a very sensitive hardware to achieve such a null test. Do you think it would be possible to achieve this ?
Old 25th January 2020
  #715
Lives for gear
 

Fourier Series Animation.
https://www.youtube.com/watch?v=ds0cmAV-Yek
Old 26th January 2020
  #716
Lives for gear
Exclamation

Quote:
Originally Posted by Tom Barnaby View Post
This should be done comparing two digital recordings, one at 48 kHz and the other at 96 kHz. It should involve actual instruments recordings made in a physical room like a piano or a voice.
Comparing waveforms froms a synthesizer would not be relevant as using 96 kHz is mainly about capturing reverberated signals in a more accurate manner.

Variations of the signal in phase would be limited to the sampling period, so we would need a very sensitive hardware to achieve such a null test. Do you think it would be possible to achieve this ?
A caveat here is that I feel all of this have already been done, though perhaps not with the particular audio interface you might have in mind. It would seem prudent to approach this problem space reading up on previous experiments and understand them on a detailed level. Generally I'd think that testing recording would be an easier prospect as nulling would reasonably then take place in the digital domain (with the recorded files) at the same sample rate.

[My own details for a suggested test removed for now]

I'd like to point out that recording in higher sample rates has good support among at least a sizeable portion of audio engineers, even if the delivery rate is 44.1. Nevertheless I suppose such a test would be interesting to learn the results from, provided detailed descriptions, faultless and documented setup as well as calibration and logically expressed objectives and result presentations. Feels like a doctoral dissertation someone has already made, so why not repeat with a similar setup when you find it?

I do see a few issues with evaluating the results though and that is if the files do not null. I expect they will not. What is the conclusion?

How are you going to distinguish in the analysis between improved whatever vs IMD or filtering differences? I feel there must be another method to evaluate the differences that are likely present. Maybe nulling is not the proper tool for evaluating what you want to evaluate here.

To me, nulling in the analog domain is a good tool for testing reproduction (DACs), which is the end where we started this exchange. Because either you can reproduce identical analog signals from the same digital sources or you cannot.
Old 26th January 2020
  #717
Lives for gear
 
Monkey Man's Avatar
 

Also, I don't see how using a single interface model would be fair because AFAIK each has its own SR sweet-spot, which technically would favour one rate over the other.
Old 26th January 2020
  #718
Lives for gear
 

Quote:
Originally Posted by Monkey Man View Post
Also, I don't see how using a single interface model would be fair because AFAIK each has its own SR sweet-spot, which technically would favour one rate over the other.
Exactly.
Old 26th January 2020
  #719
Quote:
Originally Posted by studio96 View Post
2) There is no filter giving this "proper band limiting" you are demanding. Neither digital nor analog filters can do this. There also is no known filter totally having a Zero in the stop band at all, so it is not possible to practically reconstruct everything perfectly. It is always imperfect. And the amount of "imperfectness" mostly has to do with the frequency to be reconstructed.

Or in short words: Theory is right, but you cannot reach it.
Allow me to insist on this point. Of course there is a filter giving us this proper band limiting that "we are demanding". It's not that difficult.

Let's take a closer look at what "we are demanding"!

We are demanding two things:

- A specific noise floor (in form of the right bit-depth). In case of 24bit, provably covering far more than we can ever perceive with our ears, about -144 dB.

- We want to cover the full audible spectrum + the transition-band of the filter (or maybe/arguably just half of it). An important aspect being to keep ripple, phase distortion and stop-band attenuation outside the audible region. But there are solution to all of these: Filters that produce ripple only in the stop-band, oversampling, linear phase, available spectral headroom beyond the audible bandwidth.

It is very easy to produce a filter fitting into these specs. Need the equivalent of 1500dB/oct with an error below -144dB? It's absolutely no problem.

As already mentioned somewhere above, if you want infinite bandwidth, infinitely low noise, you can get it. Given you can deliver infinite resources . There is no event or object in the universe that has these properties, though, making it a rather unproductive philosophical effort. No need to sample infinitely fast movements, they don't exist anyway. Stop insisting on sampling/capturing things that can't exist in the first place. It's exactly the impossibility of instant movement that the sampling theorem exploits, with stunning elegance and impressive practicality!

Last edited by FabienTDR; 26th January 2020 at 11:45 AM..
Old 26th January 2020
  #720
Lives for gear
Quote:
Originally Posted by Monkey Man View Post
Also, I don't see how using a single interface model would be fair because AFAIK each has its own SR sweet-spot, which technically would favour one rate over the other.
This is only interesting if one seeks the final answer of what sample rate to use. I subscribe to the notion that you cannot know what you do not test. You can only believe or assume. Possibly on good ground. I believe there is no final answer that'd be applicable everywhere.

So if you could usefully evaluate a recording test at different sample rates with one audio interface model, you would then gain the knowledge what works best for that model, wouldn't you?
Topic:
Post Reply

Welcome to the Gearslutz Pro Audio Community!

Registration benefits include:
  • The ability to reply to and create new discussions
  • Access to members-only giveaways & competitions
  • Interact with VIP industry experts in our guest Q&As
  • Access to members-only sub forum discussions
  • Access to members-only Chat Room
  • Get INSTANT ACCESS to the world's best private pro audio Classifieds for only USD $20/year
  • Promote your eBay auctions and Reverb.com listings for free
  • Remove this message!
You need an account to post a reply. Create a username and password below and an account will be created and your post entered.


 
 
Slide to join now Processing…
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Forum Jump
Forum Jump