Is realtime oversampling a dirty/destructive process?

[bmanic]

No oversampling

4x Oversampling of all compressors


Nobody commented on these. Am I imagining things or can somebody hear the annoyance thing happening at the upper mid frequencies on the non-oversampled one? Would be very interesting to hear your thoughts.

Cheers!
bManic

[Fred_Abstract]

EDIT: I realize this doesn't show the difference between low oversampling and high oversampling but the differences are very similar. I'll see if I can create a similar example with The Glue.

Cheers!
bManic

but it will show the difference between low and high oversampling in this specific plugin , probably more factor than just the oversampling are involved

[brok landers]

I thought, that's what compressors are for.

to me a compressor is an "auto gain", but if the gain envelope is going at such a high rate that it enters the audible, it's not just a gain anymore, it colors the sound. Which may not be a problem, but I wouldn't go as far as saying that it's what compressors are for.

yes, but the signal into the compressor might be continuous longer than the actual times set up, so if you set a hard attack, as soon as the release has passed and the signal runs over the threshold again, the attack cuts the signal hard (if the env follower isn't slow, which in case of a hard fast attack isn't what one wants)... so in a way you'll often end up in audio rate changes, even if not continuously. not to talk about fast attacks combined with fast releases... that's why there's often ovesampling in compressors even if they do not per se provide any audio rate modulation such as saturation.
try a clean but fast compressor (fastest release) with a smooth rhodes chord that has a chorus before the compressor (for the gain change in order to oberrun the threshold again within a continuous signal), or a phaser, you'll hear what i mean then... sometimes you'd want it to be funky with a hard attack, while other notes are held on the same track, kinda shortly after another, then you'll run into such problems, even with a very soft knee, as soon as you use fast attacks in order to get rid of some too hard transients...
just sayin'...

[bmanic]

EDIT: I realize this doesn't show the difference between low oversampling and high oversampling but the differences are very similar. I'll see if I can create a similar example with The Glue.

Cheers!
bManic

but it will show the difference between low and high oversampling in this specific plugin , probably more factor than just the oversampling are involved , making it hard to know the reason you enjoy it more

True, but it's a very similar type of thing that happens when oversampling with almost any plugin. There are things to be gained and things to loose. The best thing about oversampling in my opinion is to get rid of the annoying buildup of harshness (although it's quite subtle). It's something that I've unfortunately learned to hear and now I'm quite sensitive to it.

Cheers!
bManic

[brok landers]

But in practice, the process of filtering "whatever is above" will ALWAYS affect "whatever is below" as well.

how?

Beside obvious things such as filter passband ripple, the smooth transition region, and maybe phase shift, don't forget the Gibbs phenomenon. The latter isn't audible, but its time-domain implications can turn audible (e.g. via overloads).

It makes no sense to over generalize these things. Reducing the bandwidth introduce side-effects as a matter of fact. No matter if you like them or not.

yes. but that's only if theres only 2x oversampling, correct? isn't it sufficient enough to, say, run the whole plugin, say a compressor at double samplerate entirely and oversample only the crucial parts to 4 (or higher, if needed) times? that way you'd be able to place the filter way above human earing (as nyquist is nyquist at very hig frequency), so you should not be able to notice the sideeffects within 20hz-22,5khz?
what i'm about is, that if you can place the filter beyond the range of the human earing, it doesn't matter... same goes for aliasing... if you have a lot of aliasing, but it is at -150db, who cares, given that you use a normal levelled signal? the signal to noise ratio is the thing to look at, not the aliasing per se... just an analogy...

[FabienTDR]

yes. but that's only if theres only 2x oversampling, correct? isn't it sufficient enough to, say, run the whole plugin, say a compressor at double samplerate entirely and oversample only the crucial parts to 4 (or higher, if needed) times? that way you'd be able to place the filter way above human earing (as nyquist is nyquist at very hig frequency), so you should not be able to notice the sideeffects within 20hz-22,5khz?
what i'm about is, that if you can place the filter beyond the range of the human earing, it doesn't matter... same goes for aliasing... if you have a lot of aliasing, but it is at -150db, who cares, given that you use a normal levelled signal? the signal to noise ratio is the thing to look at, not the aliasing per se... just an analogy...

Yes, definitely. IMHO, a unified sample-rate standard around 80kHz would be optimal and make ADA converter and software dev much easier (which consequently results in better & faster tools).

[brok landers]

>snip< The best thing about oversampling in my opinion is to get rid of the annoying buildup of harshness (although it's quite subtle). It's something that I've unfortunately learned to hear and now I'm quite sensitive to it.

Cheers!
bManic

and again we're on par with our way to think... imo the most crucial thing in the digital realm is the add up between 5-10khz, where the harshness really gets annoying if you don't watch out... it's often not really noticable in single channels, but it summs up drastically on the master, especially when a clipper is overused for to hunt after the loudness in the loudness war... this is why i don't use clippers in individual channels, but only at mastering stage. only limiters with release in single channels. this really solves a lot, along with using a good sweetening eq for the high end in every single channel, to get rid of the harshness way before mastering...

[brok landers]

yes. but that's only if theres only 2x oversampling, correct? isn't it sufficient enough to, say, run the whole plugin, say a compressor at double samplerate entirely and oversample only the crucial parts to 4 (or higher, if needed) times? that way you'd be able to place the filter way above human earing (as nyquist is nyquist at very hig frequency), so you should not be able to notice the sideeffects within 20hz-22,5khz?
what i'm about is, that if you can place the filter beyond the range of the human earing, it doesn't matter... same goes for aliasing... if you have a lot of aliasing, but it is at -150db, who cares, given that you use a normal levelled signal? the signal to noise ratio is the thing to look at, not the aliasing per se... just an analogy...

Yes, definitely. IMHO, a unified sample-rate standard around 80kHz would be optimal and make lives of converter and software devs much easier (which consequently results in better & faster tools).

haha, i'm saying this since i heard bimachine on reaktor back in the days (1998?) at 132khz - i tought my ears are fooling me, the difference in signal quality was huuuughe... but so was the cpu, shooting through the roof...
yes, 80khz samplig rate for every consumer hardware, as a standard like 44.1khz is now... that would be it... would make life a lot easier...

[Nokenoku]

I can't quite put my finger on it but it's a similar constant annoyance like when a client delivers some song for mastering and he/she hasn't used any dithering when it's required but instead has truncated everything. It's a nasty harshness, usually manifesting itself from about 2kHz onward. Especially noticeable once you start hitting the final mastering limiter/ADDA loop hard.

Have you ever verified that impression in a blindtest?

Many many times. I do ABX tests all the time. A shootout between 3x tracks of The Glue set to 64x oversampling versus the same 3 tracks set to no oversampling is very easy to pick out.

The question was on dithering, not on oversampling and specific plugins.

1. Loudness of unrelated signals won't add up indefinetly.
2. The loudness of the actual audio signals (100dB louder than your "nasty stuff") will add up as well. So the relative loudness of your errors stays about the same.

.. yes, that's what I thought at first too but it seems that inharmonic stuff doesn't scale together with harmonic stuff the same way. Think about an out of tune choir. If 3 of the 100 people who sing are badly out of tune, even if they sing quietly compared to the others, something will sound off and the endproduct feels "messy". It will lack tightness and coherence.

I just made the test for you:

The loudness of quantization noise (which sounds similar to just white noise) will add about 30dB in level, if you sum 50 tracks.

So in your example it will rise from -100dBFS to -70dBFS, which would still be inaudible, even if your actual audio signal would not add up at all.

This whole "But digital artefacts will add up, and with xx tracks will become very audible." stuff is a very old and very wrong myth.
This always pops up in threads about 16bit vs. 24bit vs. 32bit vs. 64bit.

[meloco_go]

I just made the test for you:

The loudness of quantization noise (which sounds similar to just white noise) will add about 30dB in level, if you sum 50 tracks.

So in your example it will rise from -100dBFS to -70dBFS, which would still be inaudible, even if your actual audio signal would not add up at all.

This whole "But digital artefacts will add up, and with xx tracks will become very audible." stuff is a very old and very wrong myth.
This always pops up in threads about 16bit vs. 24bit vs. 32bit vs. 64bit.

I think it is incorrect to quantify audibility of errors that way. Good sounding mixes are cohesive and yet the most important things are easility discernible. In a dense mix a part that is just a little higher than the background might be noticeable and emotionally involving, but if the fidelity is compromised in bad way (there are definitely tools that reduce fidelity technically but enhance emotional responce) the part might get lost. There are definitely a threshold effects in human hearing, so small difference in noise may alter the perception drastically. It's not simply additive.

[Fred_Abstract]

what is strange some old va that probably alias like hell can still sound great and sometimes better than the cleanest plugs out there.

old virus,nords ect are still the favorites of many producers ,some with excepetional production, if aliasing was so bad it s strange? there is not a bit of hype used by plugins developers? like aliasing is really evil ect.. just wondering.

[Nokenoku]

Yes, it's not simply additive as I just explained.
It's "less than additive".

Also I don't know, what you're talking about.
Signals 70dB buried UNDER a (musical) signal are inaudible. Always.

Watch this:

[Crackbaby]

what is strange some old va that probably alias like hell can still sound great and sometimes better than the cleanest plugs out there.

old virus,nords ect are still the favorites of many producers ,some with excepetional production, if aliasing was so bad it s strange? there is not a bit of hype used by plugins developers? like aliasing is really evil ect..

I don't think this thread is about synths.

[Fred_Abstract]

but it s the same in synth , some synth can oversample the oscillator ect a crazy amount of time

[meloco_go]

Yes, it's not simply additive as I just explained.
It's "less than additive".

You don't want to hear. A signal 1dB above or 1dB below masking threshold might makes a difference between hearing and not-hearing.

Also I don't know, what you're talking about.
Signals 70dB buried UNDER a (musical) signal are inaudible. Always.

Let's not bring that here. It's too much to debate.
The point is not about being "audible noise" it is about noise that alters responce. Would you argue that dither is not needed? Tell that to J.J.Johnston featured in that video.

Also, those -70dB you say is plain lie beause you are dealing with spectrum average. When you talk about instantaneous error it might be much, much higer.
Do a simple test, get a transient rich mix and compare how something like C1 works at 44.1 and 96. You can upsample 44.1 to 96 and do a null test. It's much higher than -70dB.