48 kHZ or 44.1 kHZ ?

[poonna]

Sonic difference between 48kHz render and downsampled 44.1kHz one is non-audible. Difference between 48kHz original render and 44.1kHz original render is more obvious

48k render then downsampled to 44k still has the benefits of the processing the render at 48k, which are... higher nyquist = less EQ warping + higher automatic low passing on plugins that do it + more space for harmonics to breathe = lower aliasing.

Please stop talking nonsense. Dou you EVEN know what alisasing is?

THERE IS NO ALIASING NOW - DAWs HAVE ANTI-ALIASING FILTERS. PLUG-INS HAVE ANTI-ALIASING FILTERS. THE "LOW-PASS" YOU MENTION "IS" THE ANTI-ALIASING FILTER.

What DAW do you use BTW? And do you have measures supporting your statements, or is it "just a feeling"?

Here, take a look of what is an anti-aliasing filter, what is oversampling, and stop talking rubbish: https://www.analog.com/en/analog-dialog ... rters.html

This is probably a misunderstanding. We are not talking about the DAW's input/output or aliasing that happens in the process of oversampling, because there's no aliasing there. We are talking about the aliasing that happens because of the non-linearities in plugins. Plugins that introduce non-linearities often use oversampling to reduce aliasing and use an anti-aliasing filter on their way back to original sample rate to prevent aliasing from downsampling. There are *two* aliasing sources here: (1) the one from non-linearities which can be reduced but not eliminated by oversampling, and (2) the one from downsampling which can be prevented by the anti-aliasing filter. I believe you are talking about (2), while MogwaiBoy means (1).

Using higher sample rate can further reduce aliasing from (1) a bit by having higher Nyquist frequency so the aliasing frequencies folded back into audible range are less strong. Also, the anti-aliasing filter can have a higher cutoff frequency, so it affect the audible frequency range less. That's the point, I believe.

[poonna]

*double post*

[fmr]

This is probably a misunderstanding. We are not talking about the DAW's input/output or aliasing that happens in the process of oversampling, because there's no aliasing there. We are talking about the aliasing that happens because of the non-linearities in plugins. Plugins that introduce non-linearities often use oversampling to reduce aliasing and use an anti-aliasing filter on their way back to original sample rate to prevent aliasing from downsampling. There are *two* aliasing sources here: (1) the one from non-linearities which can be reduced but not eliminated by oversampling, and (2) the one from downsampling which can be prevented by the anti-aliasing filter. I believe you are talking about (2), while MogwaiBoy means (1).

I also have plug-ins (processors) that introduce those "non-linearities" and virtual instruments that also produce those "non-linearities". In either case I don't experience aliasing (at least that I noticed).

As you said yourself: "Plugins that introduce non-linearities often use oversampling to reduce aliasing and use an anti-aliasing filter on their way back to original sample rate to prevent aliasing from downsampling". So, hw come you end up with aliasing AFTER the anti-aliasing filter. The filter is the last stage of the processing chain. After that, the audio enters the DAW mixing chain at the project sample-rate. Where does the aliasing happen?

Using higher sample rate can further reduce aliasing from (1) a bit by having higher Nyquist frequency so the aliasing frequencies folded back into audible range are less strong. Also, the anti-aliasing filter can have a higher cutoff frequency, so it affect the audible frequency range less. That's the point, I believe.

This is undisputable. Higher sample-rate reduces aliasing by itself (when we don't have another way). But now we have (should have, at least) anti-aliasing filters (VERY GOOD ANTI-ALIASING FILTERS). Plug-ins and DAWs implement those. If a DAW has bad anti-aliasing filters, please write that, and write to them. Don't pretend that the problem is generic and affects everybody - IT DOESN'T.

And working at 48 kHz will only be MARGINALLY BETTER. The Nyquist frequency is 24 kHz, which means the safe band will be up to around 20 kHz while working at 44.1 kHz will produce a safe band up to around 18-19 kHz. If that extra bandwidth is really where you have the offending frequencies, it may be good. But what instruments have partials about 18 kHz? Cymbals, maybe?

Yes, when we are EQing, we are messing with those upper partials, and their pahses, and this may cause problems. But good EQs now have... TA-DA!... Oversampling. That's the answer to prevent those problems.

[poonna]

This is probably a misunderstanding. We are not talking about the DAW's input/output or aliasing that happens in the process of oversampling, because there's no aliasing there. We are talking about the aliasing that happens because of the non-linearities in plugins. Plugins that introduce non-linearities often use oversampling to reduce aliasing and use an anti-aliasing filter on their way back to original sample rate to prevent aliasing from downsampling. There are *two* aliasing sources here: (1) the one from non-linearities which can be reduced but not eliminated by oversampling, and (2) the one from downsampling which can be prevented by the anti-aliasing filter. I believe you are talking about (2), while MogwaiBoy means (1).

I also have plug-ins (processors) that introduce those "non-linearities" and virtual instruments that also produce those "non-linearities". In either case I don't experience aliasing (at least that I noticed).

As you said yourself: "Plugins that introduce non-linearities often use oversampling to reduce aliasing and use an anti-aliasing filter on their way back to original sample rate to prevent aliasing from downsampling". So, hw come you end up with aliasing AFTER the anti-aliasing filter. The filter is the last stage of the processing chain. After that, the audio enters the DAW mixing chain at the project sample-rate. Where does the aliasing happen?

I think you might have misunderstood the role of the anti-aliasing filter. It is to prevent aliasing from happening at the downsampling process, but it cannot eliminate aliasing that already happens. As I said, there are two sources of aliasing at work here.

(1) Aliasing from non-linear processes, such as clipping, which can be reduced by oversampling but cannot be eliminated. It happens before the anti-aliasing filter, and cannot be removed by it.
(2) Aliasing from the downsampling step of oversampling, e.g., from 96 kHz to 48 kHz, which can be prevented by using an anti-aliasing filter to remove contents above 48 kHz before downsampling. This is required, as otherwise all the contents above 48 kHz will turn to aliasing back into the 48 kHz range.

Once aliasing happens, it cannot be eliminated, as it is already mixed into the normal frequency range and cannot be taken out. Therefore, the aliasing (2) can be prevented before it happens by the anti-aliasing filter, but the aliasing (1) cannot be removed as it happens before the anti-aliasing filter. You cannot put an anti-aliasing filter before (1) happens either, as it happens in the non-linear process itself.

You don't hear aliasing in your non-linear plugins and virtual instruments probably because they do it well enough to get aliasing from (1) as low as possible (using high level of oversampling). It cannot be eliminated, but sometimes you can get it low enough to not be audible. It also depends on how much non-linear the process is. The more distortion it produces, the higher level of oversampling might be required, sometimes as high as 64x or 128x or even higher.

Using higher sample rate can further reduce aliasing from (1) a bit by having higher Nyquist frequency so the aliasing frequencies folded back into audible range are less strong. Also, the anti-aliasing filter can have a higher cutoff frequency, so it affect the audible frequency range less. That's the point, I believe.

This is undisputable. Higher sample-rate reduces aliasing by itself (when we don't have another way). But now we have (should have, at least) anti-aliasing filters (VERY GOOD ANTI-ALIASING FILTERS). Plug-ins and DAWs implement those. If a DAW has bad anti-aliasing filters, please write that, and write to them. Don't pretend that the problem is generic and affects everybody - IT DOESN'T.

And working at 48 kHz will only be MARGINALLY BETTER. The Nyquist frequency is 24 kHz, which means the safe band will be up to around 20 kHz while working at 44.1 kHz will produce a safe band up to around 18-19 kHz. If that extra bandwidth is really where you have the offending frequencies, it may be good. But what instruments have partials about 18 kHz? Cymbals, maybe?

Yes, when we are EQing, we are messing with those upper partials, and their pahses, and this may cause problems. But good EQs now have... TA-DA!... Oversampling. That's the answer to prevent those problems.

Aliasing still is one of the major issues in DSP. It is not a solved problem. The anti-aliasing filter is not the solution for non-linear process. Its use is necessary in one of the oversampling steps (the downsampling step), but it cannot eliminate aliasing from non-linear processes.

Oversampling itself is a way to reduce, but not eliminate, aliasing in the "non-linear process". Anti-aliasing filter is to prevent another aliasing from happening in the "downsampling process". They are different things.

And anti-aliasing filters have very steep cutoff. So, when it cuts off at 18 kHz, the contents above that (which there are many, not only cymbals), are gone. People who are sensitive up there can probably feel that the sound is duller.

[fmr]

Talk, talk talk... Give me proofs, not talk. And what kind of aliasing do you have with 8 x oversampling? Do you have any sources where this has been measured? It would be easy if you have a project at 96 kHz and process with one of those you pretend cause the alleged aliasing. Send me a file, with the exact description of what you did, for me to check.

The anti-alias filter doesn't cutoff at 18 kHz. It starts the roll off at around 19 kHz, which is different. The cutoff frequency is placed above 20 kHz (a little below the Nyquist frequency, which is 22.05 kHz).

Regarding the partial content (not only cymbals) point me to some spectra analysis where I can see that content. I am curious to see which instruments are those, and which intensity those partials (above 20 kHz) have.

Regarding people sensitivity to high frequencies, I know some people pretend to have bat ears, yet I still need to have proof of that too, especially when I see some timbral choices those people sometimes do.

The audio world is full of myths. It was always like this, for a long time. Unless you give me some facts to examine, I will not add anything further to the discussion. As I said, if people feel more comfortable working at 48 kHz, and they think their work is better that way, I have nothing against. Sometimes, the psychological factor has much or more important than the objective reality.

[Ploki]

Also speakers ability to reproduce those high frequencies.

anyway, since filters do screw up phase or introduce pre-ring if linear, one might argue that 48k gives filters more time to breath, moving the cutoff frequency above hearing range and allowing a little less agressive filtering.

however, all that is nulled as soon as you resample to 44.1k.
so this could only hold true if you assume that your interface has worse filters than for example iZotope does.
But if you have such a shoddy interface I question your quality priorities anyway.

[PurpleSunray]

@fmr: You really run oversampling EQs?
Don't have a single one.
Most EQ don't use oversampling because they are linear up to near nyquist and solve/reduce non-linearity by design rather than by oversample.

But back to topic, try this:
Put your non-oversample EQ (like a Pro-Q) to M+S mode. Than do a big +15db peak at 14-15khz on side-band and play a drum track on it (this is how they make "ultra-wide drums" on EDM nowadays).
I swear a can hear difference rendering this at 44.1 vs 48kHZ.. 48 sounds more "open".
Some of our DSP / EQ design gurus might be able to explain details (I never did an EQ design, so can't): huch much effect does the rolloff / non-linearity componensation / however you call it have, when I do a hugh peak at end of audible spectrum (~14-16kHZ?)

Talk, talk talk... Give me proofs, not talk. And what kind of aliasing do you have with 8 x oversampling? Do you have any sources where this has been measured?

viewtopic.php?f=33&t=511566&hilit=oversampling+EQ
he tested 4x i think

[Ploki]

Put your non-oversample EQ (like a Pro-Q) to M+S mode. Than do a big +15db peak at 14-15khz on side-band and play a drum track on it (this is how they make "ultra-wide drums" on EDM nowadays).
I swear a can hear difference rendering this at 44.1 vs 48kHZ.. 48 sound more "open".

I wouldn't do that without a multiband phase correlation meter. out of phase crap can be annoying

more "open" is a very typical audiophile expression. i hate those.

However it makes no sense to oversample an EQ. Why would you oversample an EQ? (unless it's a "colored" (aka distorted) design obviously.
edit: or dynamic eq. Pro-Q3 should have oversampling.

[fmr]

@fmr: You really run oversampling EQs?
Don't have a single one.
Most EQ don't use oversampling because they are linear up to near nyquist and solve/reduce non-linearity by design rather than by oversample.

Linear phase EQs don't need oversampling because, in theory, they don't mess up with phases, neither do they add higher frequencies.

Regarding oversampling EQs, read this, for example: https://www.joshuacasper.com/ableton-tu ... pling-eq8/

[poonna]

Talk, talk talk... Give me proofs, not talk. And what kind of aliasing do you have with 8 x oversampling? Do you have any sources where this has been measured? It would be easy if you have a project at 96 kHz and process with one of those you pretend cause the alleged aliasing. Send me a file, with the exact description of what you did, for me to check.

The anti-alias filter doesn't cutoff at 18 kHz. It starts the roll off at around 19 kHz, which is different. The cutoff frequency is placed above 20 kHz (a little below the Nyquist frequency, which is 22.05 kHz).

Regarding the partial content (not only cymbals) point me to some spectra analysis where I can see that content. I am curious to see which instruments are those, and which intensity those partials have.

Regarding people sensitivity to high frequencies, I know some people pretend to have bat ears, yet I still need to have proof of that too, especially when I see some timbral choices those people sometimes do.

The audio world is full of myths. It was always like this, for a long time.

I'm not trying to win. I just pointed out some technical inaccuracies in your posts (regarding aliasing and anti-aliasing filters). I believe Cytomic The Scream, for example, offer 64x oversampling option. 8x oversampling might be fine and can get the aliasing down quite a lot, but for guitar effects, this low-level aliasing will be amplified again in the next distortion effect or amp sim and become audible. Oversampling higher that 8x is useful at least in this application.

You can use something like DDMF PluginDoctor to see the aliasing in your plugins. It's there. I can provide some examples if I have time, but it's easy to see. You can also see the cutoff of the anti-aliasing filter on various plugins you have with this tool, too. It's pretty useful.

I don't have bat ears, and I don't claim to hear those frequencies. I'm debating this on the ground of technical accuracies only.

[Ploki]

@fmr: You really run oversampling EQs?
Don't have a single one.
Most EQ don't use oversampling because they are linear up to near nyquist and solve/reduce non-linearity by design rather than by oversample.

Linear phase doesn't need oversampling, because, in theory, it doesn't mess up with phases.

Regarding oversampling EQs, read this, for example: https://www.joshuacasper.com/ableton-tu ... pling-eq8/

this post doesnt do your other posts justice

[fmr]

However it makes no sense to oversample an EQ. Why would you oversample an EQ? (unless it's a "colored" (aka distorted) design obviously.
edit: or dynamic eq. Pro-Q3 should have oversampling.

FabFilter solved the problem (according to what they wrote), in a different way: "Internally, the dynamic EQ process will trigger on a band-limited version of the plugin's input, according to the frequency range the band works on."

Anyway, it seems they admit that some distortion may occur, because they also wrote: "Dynamic EQing also works in Linear Phase mode, but only for Processing Resolution settings up to High. The attack and release response will be slightly different from the normal behavior in Zero Latency and Natural Phase modes."

And also: "When using dynamic EQ in Linear Phase mode in combination with the Very High or Maximum resolution settings, you will see a warning sign next to the Processing Mode button to indicate that this is not possible."

[Ploki]

However it makes no sense to oversample an EQ. Why would you oversample an EQ? (unless it's a "colored" (aka distorted) design obviously.
edit: or dynamic eq. Pro-Q3 should have oversampling.

FabFilter solved the problem (according to what they wrote), in a different way: "Internally, the dynamic EQ process will trigger on a band-limited version of the plugin's input, according to the frequency range the band works on."

ah ok, still, while using plugin doctor, aliasing seems to jump to -92dB compared to zero latency nondynamic mode where aliasing is well below 170dB.

[fmr]

However it makes no sense to oversample an EQ. Why would you oversample an EQ? (unless it's a "colored" (aka distorted) design obviously.
edit: or dynamic eq. Pro-Q3 should have oversampling.

FabFilter solved the problem (according to what they wrote), in a different way: "Internally, the dynamic EQ process will trigger on a band-limited version of the plugin's input, according to the frequency range the band works on."

ah ok, still, while using plugin doctor, aliasing seems to jump to -92dB compared to zero latency nondynamic mode where aliasing is well below 170dB.

Maybe they should go with oversampling, as you said

Probably they will, sometimes in the future. Anyway, at -92dB, we are already in the "picky" universe, IMO.

[Ploki]

However it makes no sense to oversample an EQ. Why would you oversample an EQ? (unless it's a "colored" (aka distorted) design obviously.
edit: or dynamic eq. Pro-Q3 should have oversampling.

FabFilter solved the problem (according to what they wrote), in a different way: "Internally, the dynamic EQ process will trigger on a band-limited version of the plugin's input, according to the frequency range the band works on."

ah ok, still, while using plugin doctor, aliasing seems to jump to -92dB compared to zero latency nondynamic mode where aliasing is well below 170dB.

Maybe they should go with oversampling, as you said

Probably they will, sometimes in the future. Anyway, at -92dB, we are already in the "picky" universe, IMO.

at -172dB we are indisputably in "impossible to reproduce" territory.
at -92dB unfortunately, we are well within 16bits of dynamic range.

also, pretty sure that this shouldn't happen:
(input is -3dBFS 800hz sine)