.........

[astralprojection]

Yes, aliasing is a big concern when both mixing and mastering purely ITB.

Ive done lots of a/b testing with oversampling; and its not there for no reason, its there to be used. And if a plugin doesnt come with oversampling its a good idea to have a LP filter in your chain to reduce the aliasing that could occur. The effect of aliasing plagues soo many ITB tracks, and I suppose once you hear it you cannot unhear it.
In the production phase; i think its very minimal thing to worry about - since if it sounds good to you, then it sounds good INCLUDING the aliasing. But for a master; its very important to reduce it as much as you can because while you may not hear it clearly on its own, but when a/b with an aliased master the difference is there and can be quite obvious. But any ME would work to reduce that so I suppose that point is mostly irrelevant.
I guess the TLDR is; for production; if you have oversampling avaliable for a plugin that you think could cause aliasing - (high frequency content mostly from distortion) then use it.
and if you use a very long effect chain on a channel or bus (8+ plugins) throw in a linear phase LP filter at 48db/22k+ and render it down (if you dont wanna deal with the latency from the lin phase)

[imrae]

Two details about ultrasonic filtering that may not be obvious from the above:

You need to be running the session at a high sample rate (e.g. 96k); at 44.1k a 22k filter does nothing at all.

Filters should come before nonlinear processes; you cannot filter out aliasing, only condition the input signal to produce less aliasing.

[DJ Warmonger]

You need to be running the session at a high sample rate (e.g. 96k); at 44.1k a 22k filter does nothing at all.

Right

Filters should come before nonlinear processes; you cannot filter out aliasing, only condition the input signal to produce less aliasing.

Wrong.

Aliasing occurs when spectral content of a signal exceeds Nyquist frequency - whcih is half a sampling rate. At 44100 Hz, every component over 22050 Hz folds back to the spectrum, which is undesired.
At 88200 Hz signal won't fold back until it goes above 44100 Hz, but then you still have another 22050 Hz left before it reaches target frequency range.

Long story short - oversampling and cutting everything above target Nyquist frequency removes everything that could alias.

[astralprojection]

i run at 192kbit so thats absolutely correct.

Filters should come before nonlinear processes; you cannot filter out aliasing, only condition the input signal to produce less aliasing.

you can minimize it. but yes if the input has aliasing then you cannot filter it out after.

[imrae]

Filters should come before nonlinear processes; you cannot filter out aliasing, only condition the input signal to produce less aliasing.

Wrong.

Aliasing occurs when spectral content of a signal exceeds Nyquist frequency - whcih is half a sampling rate. At 44100 Hz, every component over 22050 Hz folds back to the spectrum, which is undesired.
At 88200 Hz signal won't fold back until it goes above 44100 Hz, but then you still have another 22050 Hz left before it reaches target frequency range.

Long story short - oversampling and cutting everything above target Nyquist frequency removes everything that could alias.

Ah, but that content between 22k and 44.1k isn't a problem until you run into another nonlinear process. It may be a mixture of harmonics and aliasing, but that doesn't really matter if it's in the inaudible frequency range. You can mix and EQ it without audible consequences. When you bounce to a 44.1k file, the resampling engine will filter it out as a matter of course.

Hence my advice: you only need to add extra filters before a wide-bandwidth signal hits another nonlinear process.

[DJ Warmonger]

When you bounce to a 44.1k file, the resampling engine will filter it out as a matter of course.

You don't seem to get it. To filter anything, you need to have it in separate frequency range. At 44100 sampling rate there's nothing to filter, as spectrum of entire signal is limited to 22050 Hz. If the surplus frequency is generated, the aliasing has already occured.

Hence my advice: you only need to add extra filters before a wide-bandwidth signal hits another nonlinear process.

Every non-linear process (which is practically any process you can imagine ) generates some extra frequencies.
There's that video on Youtube where I generate noise with FM by extreme aliasing of a megaherz signal ¯\_(ツ)_/¯

[astralprojection]

Long story short - oversampling and cutting everything above target Nyquist frequency removes everything that could alias.

thats a good summarization

[imrae]

When you bounce to a 44.1k file, the resampling engine will filter it out as a matter of course.

You don't seem to get it. To filter anything, you need to have it in separate frequency range. At 44100 sampling rate there's nothing to filter, as spectrum of entire signal is limited to 22050 Hz. If the surplus frequency is generated, the aliasing has already occured.

Perhaps I was not clear enough.

Consider that your session is running at 88.2k and you bounce to a 44.1k file. (i.e. from a high-sample-rate session to standard CD-quality .wav file.) The DAW runs the processors at 88.2k (could be a separate option for the render-time sample rate, let's assume this is not used) and obtains twice as many samples as it needs. There could be content > 22.05k in there. If we simply average the samples into a coarser series of 44.1k samples, high-frequency content would manifest as aliasing. To avoid this problem, any competent audio engine will filter the audio while it is being resampled, preventing the high-frequency content from influencing the output data.
The preferred implementation is to rebin with a sinc kernel. You could, somewhat equivalently, instead apply a steep 22k low-pass filter to the 96k data and follow it with a more naive resampling approach.

Hence my advice: you only need to add extra filters before a wide-bandwidth signal hits another nonlinear process.

Every non-linear process (which is practically any process you can imagine ) generates some extra frequencies.

Any process except for clean digital EQ, filters, delay, convolution reverb, stereo/gain manipulation or channel mixing. Clean slow compression, time/filter modulation and modulated reverb effects shouldn't add much bandwidth either.

For some people that'll be most of their processing chain.

[astralprojection]

Consider that your session is running at 88.2k and you bounce to a 44.1k file. (i.e. from a high-sample-rate session to standard CD-quality .wav file.) The DAW runs the processors at 88.2k (could be a separate option for the render-time sample rate, let's assume this is not used) and obtains twice as many samples as it needs. There could be content > 22.05k in there. If we simply average the samples into a coarser series of 44.1k samples, high-frequency content would manifest as aliasing. To avoid this problem, any competent audio engine will filter the audio while it is being resampled, preventing the high-frequency content from influencing the output data.
The preferred implementation is to rebin with a sinc kernel. You could, somewhat equivalently, instead apply a steep 22k low-pass filter to the 96k data and follow it with a more naive resampling approach.

you have just described oversampling what i think me and Warmonger is trying to say is that you may need to add manual LP filters to your chain in case your plugins dont have oversampling AND youre working in a higher resolution than Nyquist.

[imrae]

you have just described oversampling

I know. A high rate session with filters between all the plugins is equivalent to oversampling the plugins individually. My point is that you will realise all the practical benefits by only inserting filters in places that are both a) after the first nonlinear process, and b) before another nonlinear process.

In other cases you either have nothing to filter, or no reason to filter.

Maybe you both misinterpreted "Filters should come before nonlinear processes" to mean "one filter per channel"? That's not what I said.

[astralprojection]

i think we are all in general agreement here to be honest.
i can only talk about my own chain i and i do introduce a LP filter whenever i use something that doesnt have oversampling allready. since i work in 192kbit for mixing and mastering that makes alot of sense to me.
if youre in production phase and works in 44k or even 48k this may not benefit you in the slightest. but it could. better be sure than be sorry.