Is realtime oversampling a dirty/destructive process?

[DaveGamble]

Oversampling is just the "lame" way to solve those issues.

if anything is "lame" than its a typical approach of some devs that claims:
- aliasing is not hearable problem and thats why we dont offer any solution
- or - use higher SR if you can hear aliasing
- etc.etc...

What Im still pesonally missing is one good and dedicated OS plugin in VST form...WHY it doesnt exist???something like Crystal resampler plugin in Wavelab

My word... That's an AWESOME plugin idea... Just something that hosts a plugin at a higher sampleRate...

Has no-one done that yet?????!!??!!???!!

Dave.

[DaveGamble]

I don't know, what "filtering artefacts" you're talking about.
If you over/down-sample, then you will get filtering artefacts (as I said: not audible though ... at least, nobody could ever prove to actually hear them).

If you were talking about curve warping as "filter artefacts", then there are already enough plugins, which do not oversample and do not have that issue. There are other ways around it.

Edit:
"Lame" was not meant as "bad". It's just the easiest way.
It is totally fine to use oversampling, but I have huge respect for developers, who try to find other solutions than oversampling.

The's some talking at cross-purposes here.

Kvaca is (and correct me if I'm wrong) talking specifically about distortion plugins.
I'm not sure if techniques like RK4 count as oversampling or not. It's certainly not a brute-force oversample.
But if you want to generate a lot of harmonics, you'll need to go a long way to avoid aliasing.

Nokenoku, I think you, I, and most of us are talking about the usage of oversampling in a broader context, and I'm in agreement with your statements there.

Distortion is rather an edge case.

EQ has no business being oversampled, and I'm not sure there's been a voice of dissent raised on that matter.
Dynamics processing doesn't need oversampling by more than x2 in the audio path. However, I'm aware that some people (Andy Cytomic being a great example) measure great improvements in the accuracy of their sidechain circuit model by oversampling that, and that's fair enough. I suspect there's a good deal of extremely clever trickery alongside that oversampling too.

In general, the product of any two audo rate signals pointwise multiplied needs an oversample of x2. The convolution product of two signals needs no oversampling. (ohhh Hilbert spaces, with their inner and outer products)
But anything that's in the business of generating extreme harmonics is an open book. At this time, I tend to believe that things like RK4 and the more sophisticated 0dfb strategies are preferable to brute oversampling - but often the two go together.

Dave.

[brok landers]

I agree with Gol.. I don't like the zero latency oversample blue cat use nor the one for example in Mix Control. i use mixcontrol tons but never enable the oversample.

It messes with the phase, totally agreed. Ok, we agreed on something, first time ever

However, latency doesn't bother me as long as not on busses, so in my voxengo plugins I use the 11ms linear phase PS option. Hey, we at least agreed on 1 out of 2

Cheers

I actually prefer minimum phase oversampling to the linear phase variant. I don't know why the linear phase oversampling modes annoy me so much but they do.

Maybe I just like the "phase mess" of it all.

easy to understand why you like the phase smear - because it happens all over in the analog realm, and is, i cannot state that enough, a part of the "analog coloring" we all love so much in analog gear, and which is still not present enough in so called analog emulations. see, a simple transistor overdrive, just a tiny bit, f.e. the channel gain in an analog console like the old harrison consoles, or the raindirk consoles, sounds dramatically different when the phase is messed up a bit. if you keep the phase on a gain drive, often the result is kinda sticky (used this word in lack for a better description)... let alone that an analog eq mostly shifts the phase anyway to a certain degree... developed right, this actually serves the sound, which is why we sometimes prefer a certain eq more than another, and this is also why we often just find the sound of linear phase eq's boring... same actually goes for compressors and some other analog devices... phase smearing isn't per se something that has to be avoided like the devil, it actually serves our perception of sound-ideal, if done correctly... a lot of analog developers are actually selecting the components used to build the unit very careful in order to get that certain phase smear that they want...

[bmanic]

Oversampling is just the "lame" way to solve those issues.

if anything is "lame" than its a typical approach of some devs that claims:
- aliasing is not hearable problem and thats why we dont offer any solution
- or - use higher SR if you can hear aliasing
- etc.etc...

What Im still pesonally missing is one good and dedicated OS plugin in VST form...WHY it doesnt exist???something like Crystal resampler plugin in Wavelab

My word... That's an AWESOME plugin idea... Just something that hosts a plugin at a higher sampleRate...

Has no-one done that yet?????!!??!!???!!

Dave.

There has been a few free ones. None has been that good to be frank..

The problem is still that there are quite a few plugins that don't work correctly at higher sampling rates. Some work after reloading the plugin but others need more work. For example many of the guitar amp plugins need you to reload the cabinet impulses and choose the appropriate one for whatever sample rate you are running.

I'd still like an oversampling plugin that goes all the way to mega hertz territory just to test which plugins work.

It's one of the main benefits of Cytomic's The Glue compressor. It really DOES make a difference when you go into the silly 64x oversampling options. Sure, it takes a ridiculous amount of time to render at these levels but you can hear the benefits.

Could you consider allowing "silly" oversampling amounts for Compassion? At short attack and release times it too would benefit from massive oversampling amounts.

Cheers!
bManic

[bmanic]

Dynamics processing doesn't need oversampling by more than x2 in the audio path.

Not so sure about this. I've done a lot of testing of this and feel that it very much depends on many things:

1) What type of audio are you compressing? Lots of high/low frequency content?

2) How complex the gain reduction thing is.. a complex program dependency with a lot of "transient driven" short modulations creates a lot of inter modulation issues which seemingly get cured by massive oversampling. I don't quite understand why but this is what happens.

3) Aliasing and intermodulation distortion, even at seemingly "ridiculously low levels".. down to -140dBFS will add up and are still heard! The signal simply sounds much better and more coherent when these are pretty much completely eliminated.

4) It is extremely hard to properly measure a heavily program dependent dynamics processor. Running steady tones through a compressor is pretty much useless. Running diracs through it also only tell you a tiny amount of the full story. In the end one needs to evaluate the various negative side effects through hundreds, if not thousands, of carefully planned ABX tests on a massive amount of program material. Both in a solo and a mix context. What happens at the moment of a transient is very hard to measure, especially if there is any kind of "memory" effect involved in the program dependency.

I've also noticed that when a compressor is program dependent in the stereo-link "domain" you'll also get less stereo tearing at higher sample rates (explanation: Lets say the compressor is less stereo linked during short transients and more linked during longer signals.. or vice versa).

Finally, I agree that the most critical component is to oversample the actual sidechain/VCA block thing.. not sure if there is much benefit in oversampling the actual audio path too much.

You can of course avoid a lot of intermodulation distortion with careful design but that is a bit limiting in itself, forcing you to do certain design decisions.

Cheers!
bManic

[bmanic]

I agree with Gol.. I don't like the zero latency oversample blue cat use nor the one for example in Mix Control. i use mixcontrol tons but never enable the oversample.

It messes with the phase, totally agreed. Ok, we agreed on something, first time ever

However, latency doesn't bother me as long as not on busses, so in my voxengo plugins I use the 11ms linear phase PS option. Hey, we at least agreed on 1 out of 2

Cheers

I actually prefer minimum phase oversampling to the linear phase variant. I don't know why the linear phase oversampling modes annoy me so much but they do.

Maybe I just like the "phase mess" of it all.

easy to understand why you like the phase smear - because it happens all over in the analog realm, and is, i cannot state that enough, a part of the "analog coloring" we all love so much in analog gear, and which is still not present enough in so called analog emulations. see, a simple transistor overdrive, just a tiny bit, f.e. the channel gain in an analog console like the old harrison consoles, or the raindirk consoles, sounds dramatically different when the phase is messed up a bit. if you keep the phase on a gain drive, often the result is kinda sticky (used this word in lack for a better description)... let alone that an analog eq mostly shifts the phase anyway to a certain degree... developed right, this actually serves the sound, which is why we sometimes prefer a certain eq more than another, and this is also why we often just find the sound of linear phase eq's boring... same actually goes for compressors and some other analog devices... phase smearing isn't per se something that has to be avoided like the devil, it actually serves our perception of sound-ideal, if done correctly... a lot of analog developers are actually selecting the components used to build the unit very careful in order to get that certain phase smear that they want...

Agreed! And another important thing to note here is that the "phase smear" is NOT static! It's as dynamic as anything else. Have heavy compression and massive output gain compensation in an analogue compressor? The phase during the actual transient will be very different to the phase during a low-level signal which is under the threshold! I've been experimenting a lot with phase lately by modulating various allpass filters. It's very eye-opening.

I'm almost 100% certain that this is partly the reason why a lot of analogue equipment will lead to a mix that is way easier to "gel" together.. having no elements poking out too much. They all intermingle in a massive phase mess.. dynamically!

It can also be a highly annoying curse when you DO what things to stick out. In these cases it's easiest to stay digital. Actually a blend of both worlds usually results in very interesting mixes.

EDIT: and it's worth noting just how weird the phase response of various equipment may be! It can be 30 degrees off at 500hz during a certain amount of load, over a strange, non-regular area.. kind of like a very weird EQ shape. Not a bell, not a shelf.. something weird in-between.

Cheers!
bManic

[Nokenoku]

My word... That's an AWESOME plugin idea... Just something that hosts a plugin at a higher sampleRate...

Has no-one done that yet?????!!??!!???!!

There's WusikVM (only for synths I think), and Christian Budde has an oversampler-plugin on his page.

But to be honest ... why?
Plugins, which need oversampling usually have it already included (at least all the newer ones).



Thanks for your following post. Gives me some new ideas.

[Nokenoku]

3) Aliasing and intermodulation distortion, even at seemingly "ridiculously low levels".. down to -140dBFS will add up and are still heard!

How?

[tony tony chopper]

It's one of the main benefits of Cytomic's The Glue compressor. It really DOES make a difference when you go into the silly 64x oversampling options. Sure, it takes a ridiculous amount of time to render at these levels but you can hear the benefits.

what requires oversampling in a compressor?

The gain envelope applied is a very low-rate one, you don't need more than twice the samplerate in theory, and in practice you don't need it at all(?)

(a limiter might possibly use oversampling for a special case but that'd be splitting hairs)

[bmanic]

It creates a lot of annoying artifacts that usually rear their uggly head later in the mixing process, or at the very least in the mastering process.

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.

Lets say you have quite a few tracks in a production.. a typical number would be perhaps 50 tracks playing simultaneously (this is quite common for most pop/rock things with a lot of multitracked content). Now you have some compression on most tracks at some point in time. They all have annoying aliasing/nasty stuff below -100dBFS.. at some point in the song you will have some of these inharmonic peaks cancel each other out and at some points they will reinforce each other, basically doubling their volume.

The closest example I can give you is to take any mix with 32 tracks or more and do some heavy compression on each, with a plugin that is known to cause aliasing (or just check any dynamics processor yourself by sending a 997hz sine wave through them and look at the spectrum). Then mix them all together.

Now do the exact same thing but this time work at a higher samplingrate (or use the plugins built in oversampling) and then mix all that together.

Now take both mixes, hit them equally loud with a limiter of your choice and add a gentle high-shelf EQ from about 2kHz before the limiter. 2-3dB of gentle boost. Then listen very carefully to both mixes and do an ABX test.

I suspect one of them will "annoy" you more than the other. One will feel slightly more crowded, "messy" and irritating.

For sure, these are rather subtle things but they do add up in terms of how pleasurable the listening experience is.

The better your listening chain is the easier it is to hear these annoyances. So if you are mixing for a project which is intended to be played on basic laptop speakers then you don't have to worry too much.

If you have a chance to try with The Glue then I highly recommend squasing a few tracks with that one set to no oversampling, with fast attack and release. Then do the same with it set to 64x oversampling. There will be a noticeable difference in two areas. The actual compression will feel smoother.. like there are more "steps" in the gain reduction circuitry and the mix with the massive oversampling will feel less straining to the ears.. less annoying. You'll also notice that you can get away with more high frequency EQ boosts without killing your ears.

EDIT: one more thing.. remember that a track is usually sent to reverbs, delays, choruses and other auxiliary tracks as well. A common trick today is to have a parallel track/bus which you further process and squash the life out of. Then you do a massive mid-cut of the whole track and sneak it under the main mix. Some call this the "motown parallel bussing trick". Now all these distortions will double up in any of the sends, basically further multiplying the annoyance factor making it harder than ever to get a nice smooth mix at the mastering stage.

Cheers!
bManic

[FabienTDR]

It's one of the main benefits of Cytomic's The Glue compressor. It really DOES make a difference when you go into the silly 64x oversampling options. Sure, it takes a ridiculous amount of time to render at these levels but you can hear the benefits.

what requires oversampling in a compressor?

The gain envelope applied is a very low-rate one, you don't need more than twice the samplerate in theory, and in practice you don't need it at all(?)

Well, the gain envelope usually isn't that slow. And because of the heavy non-linear processing in the control circuitry, the control signal will easily reach Nyquist (and alias itself like crazy). This is not the worst case, it's the normal case. Of course you can slow down the control signal at a later point, but this doesn't sound that great IMHO and makes compression quite ineffective (i.e. too slow).

So, input * controlSignal will practically always create aliasing products if not oversampled (except if input's bandwidth was halfed before).

[bmanic]

It's one of the main benefits of Cytomic's The Glue compressor. It really DOES make a difference when you go into the silly 64x oversampling options. Sure, it takes a ridiculous amount of time to render at these levels but you can hear the benefits.

what requires oversampling in a compressor?

The gain envelope applied is a very low-rate one, you don't need more than twice the samplerate in theory, and in practice you don't need it at all(?)

(a limiter might possibly use oversampling for a special case but that'd be splitting hairs)

I don't know! All I know is that whenever I've heavily oversampled any of my own designs they have benefited a lot. Same with The Glue.

Also, the more complex the compressor is the more it makes a difference it seems. Like I said, it might be a case of my bad designs and that you can minimize a lot of the distortions with careful filtering/smoothing along the way but smoothing automatically slows things down which is not always desirable.

I usually don't really care when it's a question of a single track but over a whole mix these low level distortions really do seem to make a difference. So much so that I'm now a convert when it comes to doing all mixing work at 96kHz and only doing the downsampling once. The mix just "behaves" better at the mastering stage. Partly this is because I still use some rather old plugins that behave much better at higher sampling rates and don't have their own oversampling.

Why are the problems rearing their heads in the mastering stage? Perhaps it has something to do with the way a bunch of analogue equipment adds new harmonics, sometimes quite heavily so, thus also giving new harmonics to the inharmonic content close to the noise floor. The harder you push the mix for loudness, the more annoying these low-level inharmonic things get.. and you can't EQ them out.

I really do not know why or how this works.. I just know that it does make a difference to the end product. Is it absolutely _critical_ for a professional mix? No, not really. We are still talking about subtle details here but if one wants to have the ultimate listening experience then it does seem to make a difference. Of course it is very possible that I'm over sensitive to these things but I do trust my hearing.

Cheers!
bManic

[tony tony chopper]

Well, the gain envelope usually isn't that slow.

But if your compression envelope is too fast.. it gets in the audible range.. and I don't think that you want a compressor produce "sound" by itself.

We can argue that you could use a very abrupt gain change if the signal is rich or it's a transient & it will be masked, but it would only be a reason to produce the envelope at a higher samplerate, downsample it & apply it, that's still not a reason to oversample the source (although in this case oversampling 2x may be justified as there will be high freq content in the envelope).

But I still wouldn't trust a compressor in which the gain envelope is very audible by itself.

[tony tony chopper]

I really do not know why or how this works.. I just know that it does make a difference to the end product.

Ok but that means you're judging by experience with some specific plugin(s).
I have no doubt that you're hearing it and that the result is right, but it's the conclusion (that compressors need heavy oversampling) that's most likely wrong.

It's for similar reasons that people believe that a piece of music sounds better at 96khz btw. They play something at 44, something at 96, one is different or better, but that's just because the 44khz version simply isn't a bandlimited version of the 96khz one, it's just different. That, or it's more simply a case of bad samplerate conversion in a media player or whatever. I'd be careful when concluding something from testing only.

[brok landers]

>snip< Q: BTW, can you determine -3 dB change at 21 kHz?
A: Yes, it affects transients. They become "softer".
>snip<

i doubt this. you'd need a very broad slope to hear this, but the slope of a good oversampling filter is usually very steep. also, in practice there's not too many signals, especially transient rich signals, that sound good when the transient is reaching up to 21khz (in a db range that would matter), you'd usually cut it down into a frequency range where the signal actually matters (at least when f.e. going vinyl in the end)... the only exeption imo would be hihats, and even there the core frequency lies around 10-max 17khz where the main part of the signal matters...
and with any none-transient signal at 21 it would often actually benefit from the fact that it gets softer at 21khz...