You can also just listen and make up your own mind? You don’t need dbt or theorem backup for that? Just record and listen and put the resulting files in some abx program
It needs to be a blind test, because otherwise people WILL perceive the higher sample rate file to be better, just because of expectations.
a) and b) are true, but irrelevant, because it’s about mathematics and how to do the calculations, not about the result. You don’t add information by upsampling, you just allow higher frequencies to come into existence temporarily during the calculations.
And why not use it end-to-end?
Simple answer: heavy realtime process users (like modern day DAW users) pay through the nose for each processed sample in terms of CPU horsepower. As a rule of thumb, 88.2 kHz of sampling rate consumes twice the CPU cycles of 44.1 kHz sampling rate for realtime processing.
However: MORE because of memory access (caches are thrashed earlier), LESS because of upsampling and the well known limitations in parallelization in DAWs (because of busses).
c) Partially because he is part of the industry and biased? Partially because of imperfect low pass filtering in cheap converter hardware? Partially because he’s just following the cargo cult he’s used to?
I believe all three of them are true.
d) No, this argument of yours is not valid, sorry. The Nyquist-theorem fully applies here, in fact it PERFECTLY describes the analog (!) domain, even from the thermodynamic point of view (I took the opportunity during my forced absence from this forum to discuss this topic with a mathematician friend of mine - and he fully supported my point of view)
You’re mixing up two things which have nothing to do with each other - the Nyquist-theorem only explains one thing: what is the symbol rate necessary to perfectly sample a signal of a given frequency - this symbol rate is 2f.
Please don’t get me wrong, I don’t want to diss you, it’s just that you would probably get the Fields medal if you disprove the Nyquist-theorem, which I absolutely wouldn’t begrudge to you, in fact I’d LOVE to tell people that I know that cool guy who got this medal in a forum. 
What I really like about increasing CPU horsepower is the increasing quality and creativity of processing - I have used DAWs starting in the early 2000s and I remember the horrible plugins we had back then, not only their hideous UIs, but also how terrible they sounded.
Nowadays we have things like Padshop (which is stunning, people who visit me at home usually go: “wow, thats in your PC” when I show it off) or 2caudio reverbs… this is the REAL improvement we should seek when throwing higher IPC and more cores and MHz at Cubase (which I like to do).
Even with suboptimal lowpass filters in converters I don’t think we can get the same level of improvement for the same cost of clock cycles as if going from “Reverb A” (which I have seen a few days ago on a friends PC) to, say, Roomworks.
I have a strong dislike for everything esoteric, cargo culty or which is heavily advertised with flowery wording, I only buy plugins and (of course) hardware after testing them.
(Cubase, however, I buy blindly… Steinberg have never failed to impress me.)
Yes, I went with SPL for my preamp (after a lot of consideration, this was REALLY a hard decision for me) and RME for the audio interface, but really, I would NEVER go for a sample rate higher than 48 kHz - and even that only because of S/PDIF reasons, not because I like it.
This is very interesting, Patanjali:
I FULLY agree with you - and I hate the bad mouthing of MP3 a lot… a 320k MP3 is really of good quality, not distinguishable by even people with excellent hearing (only a handful of people can do that) from the already “beyond human hearing capability in reality” CD.
However, I really have problems listening to FM radio… sounds flat to me, so I avoid radio at all costs.
Again, you have done nothing to provide valid discussion, but instead provide more dismissive insults, false appeal to authority, and failed to understand what you accept as ‘truth’.
You dismiss my points a) and b) as irrelevant, but provide only a uncited vague reference to ‘mathematics and how to do the calculations’. Pathetic argument that.
You dismiss Bob Katz outright, with no citations that show why your dismissal is relevant, nor actually deal with any of his assertions. Again, argument fail.
He IS considered an expert in this field, and has shown he is willing to ‘get his hands dirty’ doing stuff about it. He willingly contributes to discussions on forums about his fields of expertise.
As for the Nyquist et al Theorem, it requires that the sampling frequency be GREATER THAN twice the highest frequency, and CANNOT be just exactly double. However, its real limitation is that it ONLY applies to INFINITELY repeated waveforms. Well, other than some electronic music, most things we are likely to record are NOT infinite, but continuously varying. So how much does that varying affect relying upon the therorem?
If you actually know something about this topic, you will know I have left you a big ‘out’ to counter the limitation I cite above. Let’s see if you can raise it!
If you are going to make statements, at least make them accurately, and know their limitations.
Also, you totally ignored my reference to the Cheung–Marks theorem, which does seem to have some relevance to this discussion. However, I would like to know how much. However, your typically arrogant, trite and shallow contributions to this topic are making you more irrelevant to it.
As for why not end-to-end, you are citing some possible economic limitations, but not engineering ones. Depending upon the material, use of end-to-end high sample rates is achievable now, as we have done it with fairly modest equipment.
This is your argument? No enumeration of the points raised, and no naming, nor citing of their relevant qualifications, of this mysterious ‘mathematician’. This is what is called ‘false appeal to authority’, and you have used it so blatantly that it thoroughly undermines your credibility. If you can’t quote, don’t pretend you can.
I want healthy discussion that can hopefully lead to relevant answers, not spurious irrelevant distractions from those incapable of allowing such to happen.
If the answer to this topic were so cut and dried it would have been dead decades ago, yet there are legitimate doubts. Dismissing those doubts, without addressing them, especially by ridicule, is NOT helping resolve this, nor the reputations of those employing such methods.
Patanjali, I didn’t mean to insult you, it’s just that I think that science should come first, before everything else.
In my day time profession I have experienced so much cargo cult related behaviour in the past, I have learned to spot it from 100 kilometers against the wind many years ago. 
Of course, Bob Katz is an expert, someone I regard highly, but this doesn’t make him an authority which overrules basic scientific facts which are known for, hm, 87 years (Nyquist-theorem was 1928, I think).
Experts tend to stick a bit too much with beloved, outdated (and I have explained where this comes from: bad filter design!) knowledge - an example would be the fact that we still discuss the impact of Hyperthreading / SMT on Cubase, which was a bad idea in the past (because of the “Replay” speculative execution system of the Pentium IV CPU), but is usually a very good thing to do right now.
But, please let me answer the very interesting objections you have brought up:
Actually, no. I have provided the example of temporarily generating very high frequencies, which is a relevant factor for, to start with, saturation plugins. When downsampling afterwards again, low pass filtering is applied to prevent aliasing, no problem there.
Another example is filtering with resonance - a zero delay resonant filter is very hard to implement (takes a lot of processing power), with extreme sample rates a few samples of delay don’t distort the signal as much - so you can go with a filter design which is actually lighter on the CPU, even though it operates on a larger dataset.
Of course there are more possibilities.
Yes, this theorem IS of (some) relevance - thats why jitter free clock generators are so popular. However, it has nothing to do with the sample rate in itself, using the Cheung-Marks theorem to explain why there might be an audible difference between 44.1 kHz and 96 kHz of sampling is akin to using general relativity to explain the operation of a grandfather clock.
As I said - I’m willing to spend €€€ on nice hard- and software, no problem there - as long as I get something in return. Cubase Pro, for example, is an amazing deal, as was my SPL Gainstation or my Stratocaster, it’s all fine.
However, the other side of economics is the returns… and I don’t see any tangible returns from using higher sample rates, it just cuts my CPU in half (more or less, I have explained the additional factors).
I don’t think the name of one of the thousands of people who studied mathematics in Vienna would be helpful to you - however, I have brought up many questions (including “doubly infinite” - his answer was: “forget that, this is a mathematical construct, it doesn’t apply here”) and he answered them all to me in a fashion which seriously underlined my point of view.
True, my fault. There is even a formula for the minimum discreete symbol rate which one can actually derive for himself.
No, it does not only apply to infinitely repeated waveforms. This would preclude the representation of information per se, the main point of the Nyquist-theorem.
Did you watch the great xiph.org video?
Marketing. Bigger, better, faster, higher, more.
Not that it matters to real life, 96 kHz just sound more and so it can be sold.
If companies were fair and marketing was truthful, they’d never started this fad in the first place.
It’s as simple as that: building a low quality 96 kHz ADC is much simpler and cheaper than building a great 44.1 kHz one - and the masses demand easily digestable numbers.
THD, SNR?
You lose about 98% of the audience with those two terms - but “96 kHz is better than 44.1 kHz” is something even an intern at a musical instruments store is able to blurt at customers.
Being dismissive of people’s concerns is disrespectful. To repeatedly do that IS insulting.
That is admirable, but so far in this discussion you have shown an inability to apply a scientific approach to your assertions and arguments.
So where was the fully enumerated steps of logical arguments that showed the chain of causality from your assertions to proving the irrelevancy of the concerns? Or at least a reference to an authoritative (that is intrinsically, and not just because some people believe it is) treatise that does that.
A self-professed ability is NOT proof of your competence, but could similarly be a proof of your incompetence, though only if there was a way to actually measure your competence in determining your own abilities! As presented, your statement has zero relevance here.
So, instead of providing actual arguments (the statement you quote of theirs was an unproven assertion from our point of view), you just want us to ‘trust you’.
So far, you have shown no substantial ability to mount rigorous arguments, nor accurately interpret what you read, so what do we really have to trust you about with respect to the veracity of the discussion you had with this associate of yours? Zilch!
Sorry, my use of ‘repeated’ was incorrect. The theorem singularly ONLY applies to doubly infinite waveforms, though not necessarily repetitive, with any time truncation giving rise to possible aliasing, thus turning what seemed so clear-cut into a probability, which usually requires some smart engineering to make up for the uncertainty. EVERY real-world usage involves time truncation.
Moreover, most applications require not waiting around until waveforms are at optimal breakpoints, so the signal gets chopped into blocks, where they don’t necessarily start and end at zero.
As I hinted at, and which you singularly failed to bring up – instead stubbornly clinging to your misconception – one method of getting a more manageable finite waveform is to add specially formulated ‘from zero ramp up’ and ‘to zero ramp down’ bands of samples to the sample block to simulate an ‘infinite’ zero waveform with a signal ‘hump’, so that the otherwise abrupt changes from chopping the signal don’t create other substantive problems.
The Nyquist et al theorem IS a theoretical construct that has its own ‘cargo cult’ following who constantly fail to appreciate its limitations. It is a useful tool that some smart engineers use to guide them in their largely empirical processes to get to what will work for their particular application. There are far too many practical limitations in any real-world sampling application to blithely rely on the solution being principally defined by the narrow application of one theorem. Their design is a whole lot of economic-engineering compromises.
You mean the one by the smug dismissive pr!ck, a self-confessed ‘snark’, who demonstrated using waveforms that are pretty well optimal for the Nyquist theorem, steady state and repetitive, conveniently the only ones they can demonstrate on the chosen equipment. I can see why you like it. Still a far cry from a violin or guitar in a room.
Ah ha. The ‘safe word’, that instantly absolves you of any need to mount any scientific or engineering sound arguments. Sorry, this is not an S&M romp, so the pain will not stop with that word.
I don’t know all the answers, which is why I am asking the questions. But I will not tolerate bogus arguments purporting to be ‘scientific’ when they amount to no more than opinion, with no reasoning shown as to why they might have any merit.
Ok, I leave it here.
Nevermind, Patanjali, it’s your CPU cycles.
44.1 kHz, because my hard-drives sound much, much better at that sampling rate. 
You don’t have to worry about Cheung–Marks, because:
but:
In every real-world digital audio application iterpolation function IS finite which makes integral of it’s square finite too. Of course finite interpolation function introduces errors in signal reconstruction, but these errors are extreamly small and most of all predictable, and can easily be made smaller than thermal noise of the electrons on your cable, if you just want to. So saying you can’t reprduce music accurately with Shannon-Nyqvist, you might as well say: you cannot transfer music accurately in any cable. True in principle, but hardly worth of second though.
Thank you very much, Jarno… I’m not a mathematician myself, I got it explained similarly by a friend of mine and couldn’t really reproduce what he said here.
What do you think, with THIS kind of background knowledge? Is there any point in higher samples, given low pass (anti aliasing) filters of very good quality?
From what I understand, not, because:
- Every waveform is just the sum of sine waves (Fourier-theorem)
- A band limited signal can bei reconstructed perfectly from samples, as long as the Nyquist criterion is fulfilled (there is only one valid, bandlimited solution)
Or am I missing something?
My friend said: no. My personal knowledge about those things says: no. The guy in the xiph.org-video says: no.
What do you say? Is there something beyond Fourier and Nyquist which DOES apply to us audio engineering guys?
Not beyond Nyquist, but because of it for some signals recorded at 44.1 (close-mic’d high volume/high frequency sources like beaded, high strings, cymbals, etc.) - high quality HPFs to keep the aliasing out of the audible range. I’d guess that is the main difference causing the large variation in SRC performances seen in those plots I referenced above. Clearly audible aliasing effects with poor SRCs under those conditions, according to some!
I wish I knew what my RME does.
I’d say being RME, if it sounds good to you, it is doing great!
I’m not either, but studied math (and computer science, physics, electric engineering, music technology, …) in 2 universities … and have forgotten most of it … but still know at least some of the basic things.
No. If both conversions and processing is done right with good anti-alias filters there is absolutely no reason to use higher sample rates. The real question is: do you have control over the whole process? Now let’s identify the problematic stages:
- A/D and D/A conversion: are you using non-oversampling converters?
1.a. If yes, definitely use higher sample rates. Analog filters are far from perfect. (OK, there’s no these kind of converters in market anymore … at least I hope so)
1.b If no, can you trust the digital filter of your oversampling converter?
I would love to see versions of plots alexis refers to, but for A/D converters. - Processing. If you process audio in the way which might introduce ultrasonic content, this content can alias into audible range if not oversampling and good anti-alias filters are used. Can you trust the writer of the plugin did his/her homework?
Yes. I record in 88.2kHz, but NOT because I think it is somehow magically superior audio format, but as an insurance against these 2 points. But then, I use my DAW just as a glorified multitrack recorder, which means I have huge amount of “free” CPU cycles to waste. If I used Cubase to create film scores with sh*tload of VSTi:s, I think I would be at 48kHz in no time.
Human hearing. Do we know enough about it? I think we do. This digital audio debate have been going on for 35 years now. If there is something more, scientists should have found it. But can we be 100% sure?
I think Steinberg could end this discussion once and for all.
Simply record in high sample rate and create a “normal” sample rate version right after (or during) recording to work with in realtime.
When rendering, Cubase then can use the higher sample rate - and life is good.
I will propose this.
…which I have done now here:
I have an old Powermac 8500 running Cubase that only runs my MIDI sequencer. I sync it to a PC running Adobe Audition to record each individual instrument to its own audio track so I can compress, limit, add reverb, etc. to each instrument track. Some of my compositions have up to 30 tracks.
I originally recorded at 44KHz 16 bits. I noticed that after recording 5 tracks that the audio quality was becoming a bit gritty. There was a loss of high frequency clarity overall. I attributed this to my inexpensive M-Audio Delta 92 sound card. But then I switched to 96kHz recording rate. Immediately the gritty tone improved in the multi track playback. When I switched to 32 bit all of the poor audio artifacts vanished. I now can record my 30 audio tracks with no tonal artifacts at all. The larger data file created by 32 bits requires a larger storage drive, but the more data bits one uses for the encoding the recording, the more detail can be accurately recorded.
I do my final mixdown to a single file and the convert that file back to 44KHz and 16 bits. My final results have improved greatly by doing my original tracks at 96KHz and 32 bit. The human ear cannot hear an individual frequency above 20KHz, but when recording multiple tracks of audio content, the blend of multiple frequencies create aliasing effects in the audible range of human hearing. I find it best to get the maximum quality of individual track recording so the end mix has the least amount of unpleasant artifacts created.
And the answer is…
96 !!!
End of story.
I’ve done blind tests listening to 44.1 vs 96 recordings. I could hear a difference, although it was VERY slight. I use 44.1 as my end product ends up there (or worse as an mp3) and the marginal gain isn’t worth the disk space being used. If I were doing something more high end, perhaps I’d use the higher freq.
-Tom
Thanks for standing up for reality! I’ve leaned lots form your series of posts and links in this thread.
Some of this “stuff” reminds me of Monster Cable hype, but our brains and ears are not able to do what they can’t do.
I’d recommend everyone to have his or her hearing checked so you at least know what you are and are not hearing. The other tremendously helpful thing for me has been to work on external monitor calibration.
Anyway, let’s all “keep clam and support reality.”
My default project settings are 48K/24 bit.
As much as I have been amazed at what some of the typical “tips and tricks” actually can do, for me, it always comes back to what I call the “Ray Charles Law” which is, “yeah, but how does it sound, baby.” I think it’s from the Tom Dowd documentary where Dowd is talking about Charles’ reaction to listening to a mix and picking out some tiny discrepancy in the levels or something like that. So after hearing someone explain all the “amazing processing they did,” the law is applied and if it passes the final Ray Charles Law test, I’m happy. http://tomdowd.com/