Should I be interested in something else than TPT/ZDF filters at this point of time?

[Richard_Synapse]

Does anyone know if any research has been done on creating (musical) filters purely from a digital non-modelling perspective?

That is, coming up with filters that sounds really good, say, for synths but not trying to mimic an electronic component based filter. I'm not talking about vanilla SVF or usual TPT, but filters with interesting character. It seems like a lot of emphasis is on modelling yesteryears analog filters.

Say, take the MS20 with its classic 'bite' or something like Steiner-Parker. Now I want to come up with filter with similar aggressiveness but from a bottom up digital implementation and for it to have its own character.

Not sure about research, but if you go back to some synths made 20-30 years ago, you may find some gems there depending on your personal preference.

Richard

[Z1202]

Does anyone know if any research has been done on creating (musical) filters purely from a digital non-modelling perspective?

The Art of VA Filter Design goes somewhat beyond the classic hardware designs (although still stays pretty close). The farthest one out is probably the 8-pole nonlinear bandpass with invertible feedback, which can produce somewhat unique sounds (there is a video demonstration). There are also generalized Butterworth filters, with nonlinear implementations by generalized SVF/SKF/Moog.

[Z1202]

Also IIRC people from Izotope looked into some other generalizations of Moog filter in their DAFX paper. And I remember reading a paper about nonlinear direct form filters, although I have to admit I'm rather sceptical about those.

[soundmodel]

Does anyone know if any research has been done on creating (musical) filters purely from a digital non-modelling perspective?

That is, coming up with filters that sounds really good, say, for synths but not trying to mimic an electronic component based filter. I'm not talking about vanilla SVF or usual TPT, but filters with interesting character. It seems like a lot of emphasis is on modelling yesteryears analog filters.

Say, take the MS20 with its classic 'bite' or something like Steiner-Parker. Now I want to come up with filter with similar aggressiveness but from a bottom up digital implementation and for it to have its own character.

This is a very good question.

I think it seems evident that analog modelling will not capture non-linearities anyways, so it could be said to be as useful to model filters based on only digital principles without considering the need to mimick some analogue circuits.

This is possibly exemplified by papers on physical modelling synthesis such as Gärder, A. (2005). Physical modeling of percussion instruments.. Such model will only capture very rough "big phenomena" in a spectral response sense. It's evident that one can approach more realism through other means than by attempting to directly model things.

[mystran]

I think it seems evident that analog modelling will not capture non-linearities anyways, so it could be said to be as useful to model filters based on only digital principles without considering the need to mimick some analogue circuits.

Exactly how is it evident? Isn't the whole point of "analog modelling" to specifically model the non-linearities?

[soundmodel]

I think it seems evident that analog modelling will not capture non-linearities anyways, so it could be said to be as useful to model filters based on only digital principles without considering the need to mimick some analogue circuits.

Exactly how is it evident? Isn't the whole point of "analog modelling" to specifically model the non-linearities?

Because all models are wrong:

https://en.wikipedia.org/wiki/All_models_are_wrong

And the papers I've read about analog models use equations, which are models about circuits.

[mystran]

Because all models are wrong:

https://en.wikipedia.org/wiki/All_models_are_wrong

https://en.wikipedia.org/wiki/Perfect_i ... my_of_good

[soundmodel]

Because all models are wrong:

https://en.wikipedia.org/wiki/All_models_are_wrong

https://en.wikipedia.org/wiki/Perfect_i ... my_of_good

Yes, but e.g. in that physical modelling paper above the results are probably not even useful, because the sounds are not musically appealing. And they can be judged for trying to recreate something that we already have (the ability to sample percussions).

I don't recall where I read it, but someone else suggested somewhere else too that musical DSP should not be driven by scientific measures, but "if it sounds good". In this sense accuracy or authenticity does not even matter.

I also thought, for example, that the Nebula convolutional libraries sound far more analog than any "modelling" plug-in I've heard. And here we see it: the other approaches it by sampling, the other by assuming the samples can be digitally recreated.

[mystran]

I also thought, for example, that the Nebula convolutional libraries sound far more analog than any "modelling" plug-in I've heard. And here we see it: the other approaches it by sampling, the other by assuming the samples can be digitally recreated.

Yet, what you are doing here is arguing for one type of model in favor of another. A sample convolution (or whatever other sampling based approach) is still a "model" just as much as a circuit simulation is a "model."

The former is usually called a "black box model" (we have a black box that we can feed with signals and observe the results, but no insight as to what is going on inside) while the latter is often called a "white box model" (when we want to differentiate from black box models; here we actually know what's inside the box) and both of these approaches are entirely valid, both of them have their own pros and cons, but most importantly they are not mutually exclusive... and realistically you always treat things as "black box" at some level anyway (eg. Ebers-Moll is a black-box model of a BJT and when you measure the parameters from a real transistor, that's a form of sampling too).

This kinda gets back to what I was trying to point out earlier though: it's not useful to limit oneself to a single method that's supposedly the "best" in all cases, rather it makes sense to figure out what method (or combination of methods) works best in each specific situation. Pretending that one method is "modelling" while another method is "not modelling" is also not terribly useful, because really it's models all the way, no matter what.

[Richard_Synapse]

I don't recall where I read it, but someone else suggested somewhere else too that musical DSP should not be driven by scientific measures, but "if it sounds good". In this sense accuracy or authenticity does not even matter.

Sure, all developers should work according to this principle. It is just that some vintage gear does sound great (not just analog gear), and hence accuracy does matter if we want to achieve a similar sound today.

Richard

[camsr]

Sometimes it's just the tuning accuracy of the filter to achieve high resonance at a desired frequency that makes these filters worth it!

[JustinJ]

What I was getting at was...imagine a world where analog synths didn't exist and we find ourselves with digital only. It's conceivable that we'd want to make, say, a low-pass filter for frequency filtering to mimic some aspect of natural sound. Our analog innovators of yesteryear came up with ladder, Sallen-Key, SEM etc to do this with different configurations giving different sound character. But would our alternate universe digital only people discover similar models? Or would they finally land on TPT and call it a day? Perhaps something novel that gives different character but that hasn't yet been discovered because of our focus on modelling analog circuits?

Just a Friday fun thought experiment.

[Z1202]

Or would they finally land on TPT and call it a day?

I can imagine one day someone would've come with trapezoidal integration or TPT equivalent, at least as a way to solve the cutoff modulation problems. Although considering how long it took to get trapezoidal/TPT back on track, even though the trapezoidal integration and zero-delay feedback techniques were effectively known for decades but went pretty much out of scope, I'm not sure how long it would have taken. Especially in the absence of analog as a reference which clearly "does better" there can be less motivation or even belief that "better sound is possible". Same with discovery of nonlinearities.

If we imagine that "analog sounds better" comes purely from the subjective effect of analog being there first, and there being no objective element to it (although I kinda doubt that), then maybe in a parallel world people will be just looking forward to all kinds of digital artifacts which in this world are rather disliked (exceptions being present).

[soundmodel]

I have to say that I am still lost as what the historical progression is in filter designs. I used to think that they advance "linearly", but after studying SVFs I come to think that maybe they're more about solution methods while the filter is sort of the same. And that some people have simply studied the wrong forms.

Some contemporary papers (http://www.dafx14.fau.de/papers/dafx14_ ... s_for_.pdf) clearly state that the attempt to apply LTI filters to musical applications is not a reasonable path at all. So I do not understand where the application of unstable and hard to use direct forms to musical uses came from. Maybe it was just an attempt to coerce filters conventional to a particular industry to use cases they weren't designed for?

Then I come to this thread where people say to discard the DFs as obsolete too.

So it just seems like it's simply reasonable to "erase" some parts of filter theory from memory and be like it never existed.

[mystran]

I have to say that I am still lost as what the historical progression is in filter designs. I used to think that they advance "linearly", but after studying SVFs I come to think that maybe they're more about solution methods while the filter is sort of the same. And that some people have simply studied the wrong forms.

Whether or not they are "the same" depends on whether you think in terms of LTI theory (transfer functions) or in terms of a more general framework.

Some contemporary papers (http://www.dafx14.fau.de/papers/dafx14_ ... s_for_.pdf) clearly state that the attempt to apply LTI filters to musical applications is not a reasonable path at all. So I do not understand where the application of unstable and hard to use direct forms to musical uses came from. Maybe it was just an attempt to coerce filters conventional to a particular industry to use cases they weren't designed for?

I'd say that's pretty much the case. Filter design in general is a bit difficult to wrap your head around, traditional digital filter design has quite an emphasis on minimizing the number of logic gates and flip-flops (which is why there is lots of work minimizing things like multipliers even if that means more adds, which is entirely pointless on a modern general-purpose FPU) and the idea that you can take a principled implicit integration method that involves solving Ax=b (ie. "inverting the system") and still typically end up with (at least reasonably) efficient code is not obvious at all except in hindsight.

There's lots of older research trying to fix the issues you end up with when you insert delays into loops to make them implementable. There's the whole WDF literature that tries to build circuits in a form where such loops don't happen. All of this seems rather pointless if we accept the idea that we can just write a bunch of differential equations and solve the thing using a principled implicit method such as trapezoidal, but the fact that this can be efficient is not obvious at all except in hindsight.

I don't think DFs are necessarily entirely obsolete as such, but given that spending an extra multiply-add (which on a modern desktop isn't necessarily even slower) for trapezoidal SVF basically frees you from having to worry too much about time-invariance and numerical precision (in musical applications where precision mostly matters at low frequencies)... I think it makes sense to take the SVF as the default and only bother with direct forms if there is some specific reason to do so... plus it's easier to learn too 'cos you can just design stuff on Laplace s-plane and not have to do a ton of trigonometry for BLT to get the coefficients... but this is not entirely obvious except in hindsight.