Oversampling

Aliasing happens the moment the harmonics are made#

A lot of people click the oversampling button without knowing why.

If a compressor is just adjusting gain over time, do you need oversampling when you move a fader? Not really. But a compressor adjusts gain very quickly, so quickly that it causes harmonic distortion. Limiting and dynamic EQ do the same. Any harmonics generated above the Nyquist frequency, half the sample rate, mirror back down into the audible range. That is aliasing, and in the worst case you can hear weird bird-chirping noises in the top end, like a low-bitrate MP3.

And you can't fix it afterwards with a low-pass filter. The aliasing happens at the point the harmonics are created. It is not like the frequencies are stored somewhere and you can get rid of them in the next step. If you have those frequencies folding back down into your audio, they are already there and you can't get rid of them later. You have to prevent it before it happens, and the way you do that is with oversampling (or ADAA).

Oversampling isn't magic#

Oversampling just puts zeros in between your samples. For 2x it pads one zero after every sample, for 4x there are three zeros:

sample, zero, zero, zero, sample.

That is called zero stuffing, and then you low-pass filter it, which smooths out all those zeros. Now you have four samples where you had one, so you have upsampled. You do your processing at the higher rate, where harmonics can go way higher without folding back into the audible range, then low-pass everything to get rid of as much ultrasonic stuff as possible, then you decimate back down to the project rate.

What needs oversampling#

Limiters, compressors, dynamic EQ and saturation should be oversampled. The need grows with shorter time constants and with any distortion built into the character of the processor. Input gain, static EQ, meters and spectrum analysers don't need it, and oversampling them just wastes CPU.

This is how Dogma's default chain works:

Input gain at the project rate, the processing chain at 4x, then the correlation ribbon and spectrum back at the normal rate.

4x or 8x is the sweet spot#

2x does a bit, but it's not really worth it. Some people say you need 32x or even 64x, but in my tests, above 8x there's rapidly diminishing returns. My free saturation plugin runs at a fixed 4x to reduce aliasing.

The cost is CPU, roughly multiplied by the factor, and it adds up over many instances. Ten years ago that was a big deal. If you compare a 2015 laptop to a current machine, there is a massive difference: at 4x you can now run way more instances than you could back then.

Oversampling makes clipping less bad, not good#

Hard digital clipping generates a huge amount of aliasing. Oversampling reduces it, so it improves things a bit. In the late 90s people bounced through second-hand DAT machines to get hotter levels that sounded better than clipping the master fader, and the reason was the good technology built into them so they overloaded gracefully, with oversampling one of the most important parts. Analogue clipping, with Zener diodes for example, has no sampling frequency and therefore no aliasing, which is a big reason it sounds better than naive digital clipping.

But even oversampled, hard clipping, to my ears, just doesn't sound that great in the first place. There are better-sounding kinds of distortion available. See clipping and saturation.

Anti-derivative anti-aliasing#

People looked for alternatives back when CPUs struggled, and one of them is anti-derivative anti-aliasing. It is interesting for simple static saturation, which is cheap to oversample anyway. When I experimented with it on extreme saturation, it started to break up and sound bad when pushed hard. Maybe my math just wasn't good enough.

The only real use I got out of it was subtle saturation with reduced aliasing.

For compression, ADAA is more or less impossible to do in any meaningful way.

There is a plugin company who claim to offer ADAA in their compressor plugin but if you read the user manual you find out that it was misleading and they aren't actually using ADAA for the application of compression itself but just for some other part of the algorithm which is a much less impressive claim.

Always be critical of marketing.