If making a impulseresponse for an amp, will that cover all about dynamics of that amp?
Or do you need many channels for a single amp?
Thinking about the differences when you pressure a tube amp to it's limits and what you get, contra running it very clean.
How much info does come from a single impuls response?
It seem to grow rapidly the number of applications for IR, microphones, amps, speakers, speakercabinets, rooms etc.
More and more are people asking for software that can run two sources and make a convolution of the two to color them.
Is what you get overrated in the sense that it does not cover so much of the true properties of a certain application.
Or is it really good?
Convolution dynamics - how does it work?
-
Christian Budde Christian Budde https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=25572
- KVRAF
- 1538 posts since 14 May, 2004 from Europe
No.lfm wrote:If making a impulseresponse for an amp, will that cover all about dynamics of that amp?
Depends on the measurement, but if you use an impulse, only the magnitude and phase of this single impulse is covered (at the fixed gain). For sweep measurements higher math is needed, which hasn't been implemented yet (in the sense of covering non-linear information properly)lfm wrote:How much info does come from a single impuls response?
These are all amplitude linear systems in contrast to an amp.lfm wrote:It seem to grow rapidly the number of applications for IR, microphones, amps, speakers, speakercabinets, rooms etc.
Valve Tone '63 of Tritone Digital is done by using only the impulse response at one amplitude. Impulse responses are interpolated in dependance of the controls.lfm wrote:Is what you get overrated in the sense that it does not cover so much of the true properties of a certain application. Or is it really good?
Regards,
Christian
-
- KVRer
- 16 posts since 25 Oct, 2000 from France
There are some mathematical methods to model non-linear systems. I had a look at these myself some time ago because i was interested in amplifier simulations, and was trying to figure out was "wrong" with current ways of modelling them.
So basically an amp is a time-invariant non-linear system, as opposed to a time-modulated device such as a wah-wah for example.
One can model linear systems such as static filters (eq), delays etc. using a decomposition of the transfer (the relation between the input and the output) function in terms of a power-series of the input/output samples.
For time invariant non-linear systems, there exists an analogous decomposition called Volterra series expansion, using a series of cross-products of powers of input samples. The existence of such a model is only theoretical, and one must devise a scheme to "extract" such a model out of measurements made from existing electronic devices. According to the bits i read, this is very difficult ...
In effect, computing the coefficients of the expansion for one particular system is far from straightforward, because for instance a careful choice must be made for the model order, to get the best tradeoff between model precision and computation costs. Stability criteria must be very complicated (didn't go as far as this), and roundoff errors must have even greater impact than in IIR linear models ...
All this to say that there exists a mathematical framework and practical tools to build proper models of non-linear devices, so there is room from progress. It seems these tools must be very difficult to implement "outside the lab", because all devices hardware (pod, sintefex chameleon ...) or software (amplitube, guitar rig) seem to employ some form of dynamic-range segmentation models, i.e one impulse response per zone of the dynamic range:
* 0 ... -10dB : ->> impulse 1
* -10 ... -20 dB ->> impulse 2, etc.
Of course there is bound to be some interpolation between overlapping zones otherwise crossing these zones would result in an abrupt change in tone.
Still, i believe this is quite unsatisfactory and that there is room for improvement.
Simulanalog team member Thomas Serafini has employed another kind of approach where each component is modelled and the engineer builds a system of non-linear equations representing the complete amp circuit. I think this tech has been incorporated in Amplitube 2. Knowing what he did with the JCM 900 emulation, Amplitube 2 might indeed live up to the hype.
So as a conclusion, maybe the way to go is to build a non-linear model based on prior knowledge of the system such as an electronic schematic, as opposed to conisdering a "black-box" approach that is possible with linear systems (deconvolution).
Cheers, i hope this isn't too obscure ..
So basically an amp is a time-invariant non-linear system, as opposed to a time-modulated device such as a wah-wah for example.
One can model linear systems such as static filters (eq), delays etc. using a decomposition of the transfer (the relation between the input and the output) function in terms of a power-series of the input/output samples.
For time invariant non-linear systems, there exists an analogous decomposition called Volterra series expansion, using a series of cross-products of powers of input samples. The existence of such a model is only theoretical, and one must devise a scheme to "extract" such a model out of measurements made from existing electronic devices. According to the bits i read, this is very difficult ...
In effect, computing the coefficients of the expansion for one particular system is far from straightforward, because for instance a careful choice must be made for the model order, to get the best tradeoff between model precision and computation costs. Stability criteria must be very complicated (didn't go as far as this), and roundoff errors must have even greater impact than in IIR linear models ...
All this to say that there exists a mathematical framework and practical tools to build proper models of non-linear devices, so there is room from progress. It seems these tools must be very difficult to implement "outside the lab", because all devices hardware (pod, sintefex chameleon ...) or software (amplitube, guitar rig) seem to employ some form of dynamic-range segmentation models, i.e one impulse response per zone of the dynamic range:
* 0 ... -10dB : ->> impulse 1
* -10 ... -20 dB ->> impulse 2, etc.
Of course there is bound to be some interpolation between overlapping zones otherwise crossing these zones would result in an abrupt change in tone.
Still, i believe this is quite unsatisfactory and that there is room for improvement.
Simulanalog team member Thomas Serafini has employed another kind of approach where each component is modelled and the engineer builds a system of non-linear equations representing the complete amp circuit. I think this tech has been incorporated in Amplitube 2. Knowing what he did with the JCM 900 emulation, Amplitube 2 might indeed live up to the hype.
So as a conclusion, maybe the way to go is to build a non-linear model based on prior knowledge of the system such as an electronic schematic, as opposed to conisdering a "black-box" approach that is possible with linear systems (deconvolution).
Cheers, i hope this isn't too obscure ..
-
- KVRAF
- Topic Starter
- 7105 posts since 22 Jan, 2005 from Sweden
Thanks to both Christian and DaPelikan.
What we might expect then is a development of this impulse stuff to get good morphing(kind of) between a number of impulses on different dB-levels, creating a mathematical model.
Maybe a single impulseresponse file with a number of noise impulses, getting analyzed and can respond more accurately to dynamics as well, is around the corner?
It is fairly easy problem to do microphones and rooms where you never stretch dynamics as much as in a guitar amplifier. The dynamic behavior of an amp for guitar is the very central characteristic sort of. You squeeze the limits in another way.
Was the work in Simulanalog done in a dynamic fashion then? Every component did have a dynamic behavior in the models, or?
Both the Marshall and Twin from Simulanalog are very good in my taste.
The Twin emulation in POD Xt was not even close to Simulanalog.
What we might expect then is a development of this impulse stuff to get good morphing(kind of) between a number of impulses on different dB-levels, creating a mathematical model.
Maybe a single impulseresponse file with a number of noise impulses, getting analyzed and can respond more accurately to dynamics as well, is around the corner?
It is fairly easy problem to do microphones and rooms where you never stretch dynamics as much as in a guitar amplifier. The dynamic behavior of an amp for guitar is the very central characteristic sort of. You squeeze the limits in another way.
Was the work in Simulanalog done in a dynamic fashion then? Every component did have a dynamic behavior in the models, or?
Both the Marshall and Twin from Simulanalog are very good in my taste.
The Twin emulation in POD Xt was not even close to Simulanalog.
-
- KVRian
- 831 posts since 7 Sep, 2004
DaPelican, on the point.
I want to add a few more thoughts, from the enigneering point of view for those who might be interested in a wider view: linear-models are by far the most important class in engineering.
Although there are non-linearities in many systems, when they reach limits, they nevertheless can be linearized (think of a curve approximated by different lines, each line representing an area, where a different model of linearization works).
So usually systems can be treated as linear (gladly, because otherwise the technological progress wouldn't have been possible).
The problem with i.e. amp-modeling is, that these units create their sound, by operating in the non-linear area of the transistors/valves.
And that is really difficult to model, because mathematically there not only need to be non-linear partial differential equations solved, but even getting a model is already very difficult.
Solving such equations is even for good mathematicans horror. Not to mention engineers...
I want to add a few more thoughts, from the enigneering point of view for those who might be interested in a wider view: linear-models are by far the most important class in engineering.
Although there are non-linearities in many systems, when they reach limits, they nevertheless can be linearized (think of a curve approximated by different lines, each line representing an area, where a different model of linearization works).
So usually systems can be treated as linear (gladly, because otherwise the technological progress wouldn't have been possible).
The problem with i.e. amp-modeling is, that these units create their sound, by operating in the non-linear area of the transistors/valves.
And that is really difficult to model, because mathematically there not only need to be non-linear partial differential equations solved, but even getting a model is already very difficult.
Solving such equations is even for good mathematicans horror. Not to mention engineers...
-
- KVRer
- 16 posts since 25 Oct, 2000 from France
@lfm: i don't know exatly how Simulanalog proceeded, but according to http://www.simulanalog.org/statevariable.pdf , quite good results can be obtained by modelling a single non-linear electronic component by a filter followed by a non-linear function. I guess one problem would then be how to combine these elements to model a complete circuit, where ther are many inter-depedent variables (signals), not taking into account aliasing introduced by the non-linearity, which can be partially reduced with over-sampling.
@Barbarossa: thanks your explanations. One "simple" way to try to solve the system is to make assumptions about what dynamic region the signal is in, and then consider the effects of the non-linearities negligible.
But as you said, in amps, especially the high-gain type, such assumptions cannot be made, because distortion is a key part of the circuit. Maybe one can still try to isolate non-linear parts of the circuit, and if the circuit can be seen as a feed-forward network of linear/non-linear elements with no feedback, then the computation can be done without resorting to very processor-heavy methods such as Newton-Raphson iteration.
I believe the expertise of the engineer is very important, because he knows what parts are really important, and where to put the processing power, and when simple linear approximations can be made. Thomas Serafini does show in http://www.simulanalog.org/tubestage.pdf that some components (here a tube stage) cannot be reduced to a filter/non-linear couple, and that many properties are lost if you do so.
@Barbarossa: thanks your explanations. One "simple" way to try to solve the system is to make assumptions about what dynamic region the signal is in, and then consider the effects of the non-linearities negligible.
But as you said, in amps, especially the high-gain type, such assumptions cannot be made, because distortion is a key part of the circuit. Maybe one can still try to isolate non-linear parts of the circuit, and if the circuit can be seen as a feed-forward network of linear/non-linear elements with no feedback, then the computation can be done without resorting to very processor-heavy methods such as Newton-Raphson iteration.
I believe the expertise of the engineer is very important, because he knows what parts are really important, and where to put the processing power, and when simple linear approximations can be made. Thomas Serafini does show in http://www.simulanalog.org/tubestage.pdf that some components (here a tube stage) cannot be reduced to a filter/non-linear couple, and that many properties are lost if you do so.
-
- KVRist
- 33 posts since 28 Feb, 2005
Hi folks,
some reading on this topic and a link to some guy that realised something like that - and patented it - so forget about reprogramming that nice algorithm ...
http://www.sintefex.com/?targ=products& ... REPLICATOR
http://www.sintefex.com/docs/appnotes/
Bye,
Tom
some reading on this topic and a link to some guy that realised something like that - and patented it - so forget about reprogramming that nice algorithm ...
http://www.sintefex.com/?targ=products& ... REPLICATOR
http://www.sintefex.com/docs/appnotes/
Bye,
Tom
-
- KVRAF
- 1743 posts since 3 Dec, 2004
that patent is most likely not as all powerful as you might think, it probably covers the exact method used in that piece of gear.TomBlock wrote:Hi folks,
some reading on this topic and a link to some guy that realised something like that - and patented it - so forget about reprogramming that nice algorithm ...
http://www.sintefex.com/?targ=products& ... REPLICATOR
http://www.sintefex.com/docs/appnotes/
Bye,
Tom
-
- KVRer
- 16 posts since 25 Oct, 2000 from France
Sintefex must have patented a particular way to implement "dynamic convolution", because the idea itself is certainly not new. In fact segmenting regions of the response of an electronic device is quite a standard method to analyze components, i learnt it at university a few years ago and i wasn't on an advance course, so it musn't be "rocket science"
And if your look a little closer at the posts above, you'll see there are several methods described, dynamic convolution being one, and maybe the most basic one, in a mathematical sense, not that i'm trying to diminish what Sintefex have achieved btw.
But maybe this discussion really belongs to the dev forum.
And if your look a little closer at the posts above, you'll see there are several methods described, dynamic convolution being one, and maybe the most basic one, in a mathematical sense, not that i'm trying to diminish what Sintefex have achieved btw.
But maybe this discussion really belongs to the dev forum.
Last edited by DaPelikan on Tue Oct 11, 2005 12:51 pm, edited 1 time in total.
-
- KVRer
- 16 posts since 25 Oct, 2000 from France
The tricky part is the interpolation, plus how to correctly design and model the different stages in the device. For example, in a compressor, you have to model the frequency-dependent characteristics (static non-linearities such as tubes and transistors) and the dynamic-response of the detector (envelope follower) + control unit (vca/opto), which aren't by essence memoryless devices ...
Surely sufficient research work for a patent.
Surely sufficient research work for a patent.
-
Christian Budde Christian Budde https://www.kvraudio.com/forum/memberlist.php?mode=viewprofile&u=25572
- KVRAF
- 1538 posts since 14 May, 2004 from Europe
In the EU it is not possible to patent a algorithm by itself, but if you patent a device then you can also patent the algorithm. That's what these guys did. Very clever. I've done some research in the field and i finally find a similar way to achieve nearly the same results. This is up to 2x faster, but this is still not enough to make it realtime applicatable without blocking half of the CPU ressources.birrbits wrote:that patent is most likely not as all powerful as you might think, it probably covers the exact method used in that piece of gear.TomBlock wrote:Hi folks,
some reading on this topic and a link to some guy that realised something like that - and patented it - so forget about reprogramming that nice algorithm ...
http://www.sintefex.com/?targ=products& ... REPLICATOR
http://www.sintefex.com/docs/appnotes/
Anyway it is a nice piece to make snapshots of analog gear in high quality. After this first step one can start developing a physical model for this. The 'classic' mode of posihfopit (equals the 'Analog' mode in 'Electri-Q') has been done with this.
With the build in Pultec-EQ filter clones i've done the same, but they turned out to be very linear, so i use a linearized model here.
Ups, don't want to write so much details...
Christian
-
- KVRian
- 831 posts since 7 Sep, 2004
Keep it rolling. Interesting!