True analog modeling?
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
A recent post about Hive got me searching for true analog modeling. U-he and Gforce are the only ones I can think of. I'm unsure about that xils lab stuff but I read that he came from arturia which makes me think it's sampled wave forms used as oscillators. Since I'm having troubles finding true analog modeling can some of you point me towards the good stuff?
-
- KVRAF
- 3897 posts since 28 Jan, 2011 from MEXICO
Xils stuff is also quite nice and I dont think it uses samples, neither Arturia which has been criticized for sing the same oscilators for all its synths so not really emulating the specific synths.
Personally I love Diva, but Monark and Lush are also very nice.
Personally I love Diva, but Monark and Lush are also very nice.
dedication to flying
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
Well, In lush you choose a sample as oscillator right from the get. It's sample based. As far as NI is concerned, Absynth is sampled waveforms, Massive obviously, FM is probably ... I never liked NI because their filters are bland and sound 2 dimentional plus I don't like the fact that you have to install the service center.etc.... BUUUUUUUT I have not heard or played with monark and it is getting a lot of attention. It could be analog modeled so I would have to find a way to try it out. It could be though and that's why every one seems to be high on it. I'll look into it.
-
- KVRAF
- 8414 posts since 4 Jul, 2012 from Alesia
Here's some interesting reads if your into this sort of thing.
https://en.wikipedia.org/wiki/SPICE
https://en.wikipedia.org/wiki/Semicondu ... e_modeling
https://en.wikipedia.org/wiki/Transistor_model
https://en.wikipedia.org/wiki/Diode_modelling
https://en.wikipedia.org/wiki/SPICE
https://en.wikipedia.org/wiki/Semicondu ... e_modeling
https://en.wikipedia.org/wiki/Transistor_model
https://en.wikipedia.org/wiki/Diode_modelling
-
Echoes in the Attic Echoes in the Attic https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=180417
- KVRAF
- 11054 posts since 12 May, 2008
I think you've got some wrong ideas here. Lush is not sample based. It is a DSP generated waveform just like Diva. Ie. computer generated impulses like an analog synth does, but digitally generated. Absynth has several oscillator types, most of which are dsp oscillators, though there is the sample or granular type which use samples. The rest don't. I wouldn't say Massive uses samples, they are wavetables which isn't exactly the same. FM8 uses DSP waveforms, no samples.Nanners wrote:Well, In lush you choose a sample as oscillator right from the get. It's sample based. As far as NI is concerned, Absynth is sampled waveforms, Massive obviously, FM is probably ... I never liked NI because their filters are bland and sound 2 dimentional plus I don't like the fact that you have to install the service center.etc.... BUUUUUUUT I have not heard or played with monark and it is getting a lot of attention. It could be analog modeled so I would have to find a way to try it out. It could be though and that's why every one seems to be high on it. I'll look into it.
The analog modelling thing is more about the signal path and the effect of that on the oscillators. The amp gain, the filters especially. The way oscillator functions like sync and fm work and so on. So lush for example models the SH101 oscillator setup, signal path and filters, as well as the response of the modulators like envelopes, which can behave differently in different synths. When it comes to the oscillators of the synth, it's subtle adjustments to the shape based on the real world thing it's trying to emulate. For example a saw might not be a perfect saw, but slightly rounded or something. IT might also be able to have a slight drift out of tune or something like that to model the real world behaviour of an analog device.
-
- KVRAF
- 8414 posts since 4 Jul, 2012 from Alesia
I think In order to truly model an analog system, it would require a very powerful computer system that could model in real-time the behavior of the electricity itself on an atomic level and how it interacts with the components. I don't think we have the technology or power available yet to even come close to that level of realism.
Of course this is my opinion and I am far from a DSP expert, but just based on what i've read.
Here is a quick overview on basic electricity for those who are interested.
http://reprise.com/host/electricity/atoms.asp
Of course this is my opinion and I am far from a DSP expert, but just based on what i've read.
Here is a quick overview on basic electricity for those who are interested.
http://reprise.com/host/electricity/atoms.asp
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
I've been waiting to see if this gets moved, But it's my understanding that in order to lower the cpu cycles the guys coding this stuff have resorted to taking a digital copy or image of a waveform and basically using that image played back at different frequencies to save the cpu from having to calculate in realtime what it looks like. I've lumped these into "sample based" as well as pure samples. It's not what I'm looking for, I want the good stuff. Ace and Diva do it right. I believe that Gforce does it right...those are the only two I can think of. I'll go check out that reading material vortex and see.
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
Echoes in the Attic wrote:I think you've got some wrong ideas here. Lush is not sample based. It is a DSP generated waveform just like Diva. Ie. computer generated impulses like an analog synth does, but digitally generated.Nanners wrote:Well, In lush you choose a sample as oscillator right from the get. It's sample based. As far as NI is concerned, Absynth is sampled waveforms, Massive obviously, FM is probably ... I never liked NI because their filters are bland and sound 2 dimentional plus I don't like the fact that you have to install the service center.etc.... BUUUUUUUT I have not heard or played with monark and it is getting a lot of attention. It could be analog modeled so I would have to find a way to try it out. It could be though and that's why every one seems to be high on it. I'll look into it.
What's all this choosing a timbre about? Looks like samples.
-
- u-he
- 28065 posts since 8 Aug, 2002 from Berlin
Well, my perspective on analogue modeling has changed over the years - mostly because I became proficient with different aspects of it, such as being able to apply nodal analysis on a circuit and transform the equations into code.
"True analogue modeling" in my current view does not require modeling things on an atomic level, but it surely requires consistency with the laws of physics and a level of detail that surpasses perception. If I say "True" I mean methods that don't just use portions of analogue modeling and then deviate from there, e.g. for performance reasons.
Creating an analogue model of, say, a filter is actually not that difficult. There's a handful of fundamental equations derived from laws such as Ohm's ( Current = Resistance x Voltage ), one can write them down as they apply to a circuit and one can solve them. This works very easy for filters that do not distort or for oscillators that do not "phase wrap". In these cases all equations can be thrown into a single equation for each voltage or current being thrown into the system and gathered at the output.
However, things become difficult as soon as elements with non-linear behaviour enter the scene. Once there's a transistor, an OpAmp or a diode the equations become a bit more complex. Most of all, they become impossible to solve as a single equation per node. They can still be solved with perfect accuracy, but this requires to calculate them several times much like a trial and error method: One makes an estimate for some unknown variables, one calculates the equations, one checks out how things add up, one takes the results to make a better estimate and does the process again - until things add up. Thats what an iterative numerical solver does, and these methods are said to be "numerical".
As is immediately clear, linear analogue models are a piece of cake for the CPU while non-linear ones require a multitude of processing power. For many applications it's good enough to just approximate the non-linear behaviour. One doesn't always have to go the full way. While one can still base an algorithm on a linear analogue model (i.e. replace transistors by resistors) there are a gazillion methods to apply those non-linear effects differently. This includes lookup tables for simple distortion, geometric approximations of curves, or plain cheating by applying the distortion afterwards.
The most elaborate of latter methods are those that combine geometric approximations and a few iterative steps to refine the result. These are often called "analytical methods". If done well the outcome is indistinguishable from the numerical ones - but they still require a good extra deal of CPU power.
#---
From above it's quite obvious that Diva (and to my best knowledge Monark) use iterative solvers in their filter models - which is the main source for their CPU hunger, but also the main source for their authenticity. The Drop on the other hand (and as far as I understand Andy) uses a most elaborate method to approximate first and then refine to perfection. I think those and probably a bunch more that I'm not familiar with is what I'd call "true" analogue models.
Hive on the other hand has filters that are based on linear analogue models (OpAmps/Transistors substituted by resistors). In the "Normal" and "Dirty" engines however I've done a great deal of cheating. I've used a secret little trick to sneak distortion into the path while maintaining a solution based on single equations. In this case, instead of iterating the equations, I calculate the distortion itself a few times. While this results in nothing like accurate analogue behavior, it still sounds highly musical, with all the benefits of analogue modeling, such as zero delay feedback, tuning accuracy and numerical stability. It takes methods of analogue modeling but it isn't a true analogue model (your CPU will thank us for that)
Cheers,
- U
"True analogue modeling" in my current view does not require modeling things on an atomic level, but it surely requires consistency with the laws of physics and a level of detail that surpasses perception. If I say "True" I mean methods that don't just use portions of analogue modeling and then deviate from there, e.g. for performance reasons.
Creating an analogue model of, say, a filter is actually not that difficult. There's a handful of fundamental equations derived from laws such as Ohm's ( Current = Resistance x Voltage ), one can write them down as they apply to a circuit and one can solve them. This works very easy for filters that do not distort or for oscillators that do not "phase wrap". In these cases all equations can be thrown into a single equation for each voltage or current being thrown into the system and gathered at the output.
However, things become difficult as soon as elements with non-linear behaviour enter the scene. Once there's a transistor, an OpAmp or a diode the equations become a bit more complex. Most of all, they become impossible to solve as a single equation per node. They can still be solved with perfect accuracy, but this requires to calculate them several times much like a trial and error method: One makes an estimate for some unknown variables, one calculates the equations, one checks out how things add up, one takes the results to make a better estimate and does the process again - until things add up. Thats what an iterative numerical solver does, and these methods are said to be "numerical".
As is immediately clear, linear analogue models are a piece of cake for the CPU while non-linear ones require a multitude of processing power. For many applications it's good enough to just approximate the non-linear behaviour. One doesn't always have to go the full way. While one can still base an algorithm on a linear analogue model (i.e. replace transistors by resistors) there are a gazillion methods to apply those non-linear effects differently. This includes lookup tables for simple distortion, geometric approximations of curves, or plain cheating by applying the distortion afterwards.
The most elaborate of latter methods are those that combine geometric approximations and a few iterative steps to refine the result. These are often called "analytical methods". If done well the outcome is indistinguishable from the numerical ones - but they still require a good extra deal of CPU power.
#---
From above it's quite obvious that Diva (and to my best knowledge Monark) use iterative solvers in their filter models - which is the main source for their CPU hunger, but also the main source for their authenticity. The Drop on the other hand (and as far as I understand Andy) uses a most elaborate method to approximate first and then refine to perfection. I think those and probably a bunch more that I'm not familiar with is what I'd call "true" analogue models.
Hive on the other hand has filters that are based on linear analogue models (OpAmps/Transistors substituted by resistors). In the "Normal" and "Dirty" engines however I've done a great deal of cheating. I've used a secret little trick to sneak distortion into the path while maintaining a solution based on single equations. In this case, instead of iterating the equations, I calculate the distortion itself a few times. While this results in nothing like accurate analogue behavior, it still sounds highly musical, with all the benefits of analogue modeling, such as zero delay feedback, tuning accuracy and numerical stability. It takes methods of analogue modeling but it isn't a true analogue model (your CPU will thank us for that)
Cheers,
- U
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
Hey Urs, can't believe you took the time to answer that. Thanks and I'm taking the time to absorb it. I understand what you are saying... it comes down to modeling physics and that's a casualty or quantum calculation that in itself can only be approximated by chance. So there are no definite answers because in the real world there is only probability. I see what you are saying. Let me absorb that for a minute. Thanks again. (now get back to work!, hahaha)
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
Yeah, you know that's the thing that I'm after, the imperfections of probability. That's the character of a good instrument. That's what I appreciate about your earlier offerings. I have that same appreciation for Gforce (rightly or wrongly, but they seem to haver "it"). I have a problem with the static and predictable (what I call two dimensional) sound of most soft synths. Maybe I was quick to dismiss Hive though. What you are saying is that in an attempt to get some of that character into an instrument you simply asked the processing to calculate certain variables that were important for the character and not all the variables.
-
- KVRist
- 52 posts since 18 Mar, 2014
This reminds me the P-code method in FEA; raising the polynomial order to solve iteratively, non-linear analysis is never perfect, but then again neither is pi.
BC
-
- KVRist
- Topic Starter
- 34 posts since 10 Jul, 2015
[quote="Battle Cat"non-linear analysis is never perfect, but then again neither is pi.[/quote]
Exactly. That made me smile.
We all have certain natural laws that we live by and take as influences I guess. One of mine is perfection is over rated.
Or in other words the Imperfection is where the beauty is.
Exactly. That made me smile.
We all have certain natural laws that we live by and take as influences I guess. One of mine is perfection is over rated.
Or in other words the Imperfection is where the beauty is.
-
- u-he
- 28065 posts since 8 Aug, 2002 from Berlin
Well, hehe, the equations have nothing to do with probability. That's the thing - once you're dealing with billions of electrons doing the same thing with a certain probability, the probability factor itself becomes the actual factor.
My stance on analogue is that it is *generally* more exact than digital, specifically in small effects that can be described as "high resolution". For example, two subsequent waveform cycles of an analogue oscillator are usually much more identical than two subsequent cycles of a digitally generated waveform - simply because the "resolution" of analogue is nearly infinite while digital stuff is chopped into sample sized frames. Similarly the capacitors in a filter charge continuously while their charge has to be approximated in a digital filter from sample to sample.
However, the imperfections of analogue are undeniable as well - resistors dissipate heat, heat dissipation changes resistance and therefore frequencies drift. Likewise, no two diodes are ever exactly the same, hence distortion curves are asymmetric. Those effects are usually a bit random, but they're also subject to slow change: At first a Jupiter 8 oscillator can't have 0% pulsewidth, but after twenty minutes it can (Your Jupiter May Vary)
My stance is that these "slow" effects can usually be observed and mimicked without too much fuzz. The real problem about analogue modelling is to match the perfection of analogue in the digital domain. One way to do so is to use estimates and probability in the code, to work out how the analogue stuff actually works. The better the estimate, the less CPU until we're there.
- U
My stance on analogue is that it is *generally* more exact than digital, specifically in small effects that can be described as "high resolution". For example, two subsequent waveform cycles of an analogue oscillator are usually much more identical than two subsequent cycles of a digitally generated waveform - simply because the "resolution" of analogue is nearly infinite while digital stuff is chopped into sample sized frames. Similarly the capacitors in a filter charge continuously while their charge has to be approximated in a digital filter from sample to sample.
However, the imperfections of analogue are undeniable as well - resistors dissipate heat, heat dissipation changes resistance and therefore frequencies drift. Likewise, no two diodes are ever exactly the same, hence distortion curves are asymmetric. Those effects are usually a bit random, but they're also subject to slow change: At first a Jupiter 8 oscillator can't have 0% pulsewidth, but after twenty minutes it can (Your Jupiter May Vary)
My stance is that these "slow" effects can usually be observed and mimicked without too much fuzz. The real problem about analogue modelling is to match the perfection of analogue in the digital domain. One way to do so is to use estimates and probability in the code, to work out how the analogue stuff actually works. The better the estimate, the less CPU until we're there.
- U
