Best sound quality - run 96khz samplerate, do not oversample.

[IncarnateX]

This is interesting, but not what we are talking about here at all.

This thread is about the best way to minimize aliasing artifacts from non-linear plugs, while the study above is only testing the final delivery format.

Oh..hmm! I just read the thread title and the opening post and these sure leave the impression that the thread about sound quality relative to samplerate, but never mind then.

[IIRs]

This is interesting, but not what we are talking about here at all.

This thread is about the best way to minimize aliasing artifacts from non-linear plugs, while the study above is only testing the final delivery format.

Oh..hmm! I just read the thread title and the opening post and these sure leave the impression that the thread about sound quality relative to samplerate, but never mind then.

Specifically it is about running a higher host samplerate as an alternative to per-plug oversampling.

[Compyfox]

This isn't really a problem of audio interfaces - almost every interface supports 96 kHz and even 192kHz. Under the hood they may even work at 384kHz and perform digital downsampling.

But it's still a drivers and datarate transfer problem. They'd be the bottleneck then. On top of having higher CPU spikes at higher sampling rates. At least IMO.


If this all could be handled automatically WITHOUT a larger CPU, ASIO or data transfer spike - I'm all for it. But you still have the final format 44/16 and 48/24 for consumers.

[khanyz]

@Urs

But wouldn't the Nyquist/2 filter between processes I mentioned remove the 40KHz?

I'm not saying keep the content that's going to cause aliasing and is not required, I'm saying remove it but skip the up/down sampling so as not get any artifacts from it. So the range between nyquist/2 and nyquist will be empty between process.

And going further with that and the 4x oversampling. Is there then a case to use the 176.4/192 band throughout in order to aid processing, but again remove anything above Nyquist/4 between processes. This would remove the need to up/down sample.

Cheers,
Nigel

[@midnight]

I think with 6-core CPUs currently being released we should be talking about working at 192kHz as a standard - which is like an "ultimate solution" for high-quality "in the box" sound processing. Going higher than 192 kHz is usually only marginally better. Working at 384kHz like Pyramix does covers all the possible bases, but it's too much I think.

Actually, regarding 192khz, Dr. Lavry thinks even that is overkill. His opinion is that the ideal sampling rate is in the 60khz.

However, it's interesting when reading his paper about it, he doesn't seem to mention aliasing *at all* ...

Is it possible that Dr. Lavry has completely overlooked aliasing in his analysis? Here's the PDF for all to read:

http://www.lavryengineering.com/documen ... Theory.pdf

A lot of this goes over my head, but I perused it and didn't see any mention of aliasing.

[Aleksey Vaneev]

@midnight, from theoretic standpoint, 44.1/16bits storage format is absolutely adequate for most consumer uses, and everything else boils down to consumer equipment that succeeds or fails reproducing the audio stored in that format.

But this topic is about production, not about delivery. And it is during production where aliasing happens. Beside that, I strongly suggest recording cymbals at 96kHz at least. Just because this way most ADCs will perform adequately. Cymbals are quite loud above 22 kHz and may alias on average ADC.

[milosh_]

you guys want to blow my mind or what ?

[@midnight]

But this topic is about production, not about delivery.

Indeed, I am the topic starter

So Lavry's paper is about digital audio storage, and not about the recording aspect of digital audio. Would this explain why he doesn't get into aliasing at all?

[Urs]

@Urs

But wouldn't the Nyquist/2 filter between processes I mentioned remove the 40KHz?

I'm not saying keep the content that's going to cause aliasing and is not required, I'm saying remove it but skip the up/down sampling so as not get any artifacts from it. So the range between nyquist/2 and nyquist will be empty between process.

And going further with that and the 4x oversampling. Is there then a case to use the 176.4/192 band throughout in order to aid processing, but again remove anything above Nyquist/4 between processes. This would remove the need to up/down sample.

Cheers,
Nigel

Yes.

However, inserting steep filters at Nyquist/2 (88/96k) or Nyquist/4 (176/192k) is essentially the same as downsampling and upsampling. In fact, when using those polyphase filters I mentioned, it's completely identical - a so called polyphase halfband filter that kills anything above Nyquist/2 does exactly the same computation as using two polyphase filters for downsampling/upsampling inbetween two effects.

(sorry for the tech talk)

Hence, it really doesn't matter if you work in 96kHz and filter inbetween processes, or if you work in 48kHz and oversample each process. The result is the same, both in outcome and cpu footprint.

Urs

[khanyz]

(sorry for the tech talk)

No, tech talk is good. It's what I understand and gets straight to the point.

It seems to me though, that there should be a difference. I'll have to go through the math and see what compounding can be done.

Looks like a futile quest I can throw myself into.

BTW: Do the filters really have to be that steep? Can you move them away from binary fractions of Nyquist, if you don't do the up/down sampling. I'll have to check the practicality of that too.

Thanks,
Nigel

[Aleatoriac]

bit-crushed bat sonar.

That is awesome and inspiring.

[Jafo]

I'll try an example (mind you, I'll use a shortcut to avoid bringing intermodulation distortion into the game):

Think of having a signal chain with two identical vintage equalisers in flat setting. Because they are vintage and blah they add the 3rd harmonic (possibly among other things), two octaves above the original frequency.

Now we have 96kHz sampling rate.

Our signal may have a loud presence of a 10 kHz signal (self oscillating synth filter?)

After the first compressor, that 10 kHz signal has got an harmonic overtone added two octaves above, at 10 x 4 = 40kHz, maybe 20dB lower. Fine. We have 48kHz bandwidth, the 40kHz are no problem.

But... we have a second equaliser. That one adds another 3rd harmonic. So our 40kHz harmonic gets its own harmonic added, at 4 x 40 = 160kHz, at audible 40dB below the original 10kHz sound.

Now, 160kHz don't fit into 48Khz. It does Nyquist/DC bounce limbo: They'll bounce back from 48kHz to -112 kHz, forth from 0Hz/DC boundary to 74kHz and back again from 48kHz down to 12kHz. (not quite sure if I got the math right here...)

I'm not sure either. IIRC, each harmonic is an integral multiplier of the harmonic order, so the third harmonic of 10k would be 30k. Subtract one from the harmonic order and you've got the overtone order -- the fourth harmonic is the third overtone, and is at 4x. At least, according to this page.

But that's just a nitpick in terminology, and the rest of your point is awfully compelling. Oversampling to 96/88.2 only gives you one more octave of clarity at the top -- helpful, but no silver bullet.

Waveshaping (most of waveshaping?) tends to generate infinite harmonics, especially ones that don't reduce in amplitude quickly enough. I honestly thought that this would be enough to model tubes, but after many, many hours with various functions said to be tubelike (the usual suspects from Pakarinen, Doidic, Araya, blah blah blah et cetera ad infinitum and nauseam), running them through SPAN and FuncShaper showed me that you just get an undynamic fuzzy mess that's going to alias within too many circumstances. Bletch.

I think that different mathematical techniques are called for to handle nonlinearities. Matrices, differential equations, and physical modeling adorn Yeh's thesis and the fabled Koren equations, as do papers about Asynth, the Simulanalog sims, Juicy 77, and the TSE sims. Time spent with SPAN shows that these generally don't generate infinite harmonics. Not to mention being much more convincing than most other sims. Now if only I understood the math(s).

Elsewhere in this thread you mention polyphase filters. Would you expand on this?

There you go. By not bandlimiting things inbetween two processes, our 10 kHz signal got a non-harmonic aliased signal at 12kHz, with just twice the most common type of harmonic distortion in series.

#---

Now, had each equaliser filtered anything above 24kHz, none of the aliasing in this example would have occurred.

Hunh, that would be a lot simpler and easier than using better math.

Another thing that goes to show is, oversampling to twice the sample rate does hardly help with the most common harmonic distortion. You have to go 4 x oversampling or 192kHz sampling rate to wipe out the most ordinary aliasing.

That too.

(that's because most non-linear processes apply symmetric waveshapers which emphasize odd harmonics. Whereas "tube like" distortion that creates even harmonics such as 2nd is rather rare)

Cheers,

Urs

More precisely, symmetric waveshaping generates only odd-order harmonics; as more distortion is applied, these generated odd-order harmonics overpower any even-order in the original signal. Same with push-pull output stages (which are seen in the vast majority of guitar amps and their associated sims).

[Aleksey Vaneev]

So Lavry's paper is about digital audio storage, and not about the recording aspect of digital audio. Would this explain why he doesn't get into aliasing at all?

Yes, I think his paper is about representation and storage only. But the paper is flawed in that during recording various technical aspects of the recording equipment come into play. And this is where aliasing may happen easily.

[tony tony chopper]

Oversampling to 96/88.2 only gives you one more octave of clarity at the top -- helpful

an octave fully inaudible, so how is it helpful? Yes, helpful but only if you plan to make it audible (by pitch shifting or resampling down..), but that's for special effects.

[khanyz]

an octave fully inaudible, so how is it helpful? Yes, helpful but only if you plan to make it audible (by pitch shifting or resampling down..), but that's for special effects.

It's helpful for processing headroom. You would remove it after you've finished processing and want to listen to it.

It's like having a couple of extra decimal places when you do calculations, then rounding up for the final result.