OP
- Thread Starter
- #581
Me too.
-
Welcome to ASR. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!
DIY 3D Speaker Scanner - the Mathematics and Everything Else
- Thread starter NTK
- Start date
mwmkravchenko
Senior Member
The secret sauce math behind Klippel was from a Thesis done by one of the students at the school Wolfgang taught at. Neither I nor another person involved with this has ever been able to find that Thesis. We know the name of the paper. And believe me, we looked for it. We never found it. So, the math derivation used presently has nothing to do with the original work done to get the Klippel system going. So no infringement on Klippels basis for the measurement system. The mechanical system is partly the same if you take a spinning axis microphone boom under CNC control as being the same. Is it the same? Not really. So many changes it is easily not an infringement.
So the two main parts of the HALS system are indeed different. And that is what sets this far away from patent infringement.
Mark
jahargreaves
New Member
- Joined
- Jun 2, 2026
- Messages
- 4
- Likes
- 17
Hello, my first time here too. All very nice work, and something I've been interested in for a long time too.
What is the name of the paper? I have looked for it too previously but never had a name. Someone told me it was an MSc by research at RWTH Aachen.
I assume you know about this paper already? (apologies I couldn't be bothered trawling through 30 pages to check)
Also I don't know if you know this already but the Klippel NFS linear positioning elements are built by a company called Igus.
All the best!
PS you can find some of my work related to this here:
- https://salford-repository.worktrib...ic-centre-concept-to-directional-loudspeakers
- https://salford-repository.worktrib...ons-including-source-and-receiver-directivity - it's not discussed here but you can apply this approach to loudspeaker directivity too
OP
- Thread Starter
- #585
Dr. Hargreaves, welcome to ASR!
I think/believe a significant chunk of the Klippel NFS development work was done by Christian Bellmann at TU Dresden (Dresden Technical University) and later at Klippel. The only mentioning to his thesis at TU Dresden I could find was from AudioXpress articles he coauthored (links below, see "About the Authors"). I, however, was not able to find a copy of his thesis online.
Practical Test & Measurement: The Proof Is in the Testing
The Klippel Near Field Scanner System (NFS) was used for the critical testing and measurement of Holoplot's X1 Matrix Array compensation algorithm capabilities. All our readers who have appreciated the story of the development of the unique Holoplot technology, published in the September 2022...
Directivity Measurement of In-Wall Loudspeakers
In this article, the authors share their research on new requirements for the measurement of in-wall installation speakers. The directional characteristics are an especially important factor, because the direction in which a loudspeaker emits sound highly affects the interaction with the...
Bellmann is also one of the authors of this early Klippel NFS presentation.
Tom Kamphuys
Member
- Joined
- Feb 15, 2024
- Messages
- 82
- Likes
- 238
Hello @jahargreaves,
I think you'll find this interesting:
github.com
I think you'll find this interesting:
lah-scanner/documentation/centre_origin_help.md at main · dfapinov/lah-scanner
Python pipeline for 3D loudspeaker measurement and sound field reconstruction. Processes impulse responses on a spatial grid, performs spherical harmonic expansion (SHE), separates internal/externa...
github.com
jahargreaves
New Member
- Joined
- Jun 2, 2026
- Messages
- 4
- Likes
- 17
Thanks for the welcome and info. Most universities by default make thesis available online these days as you probably know, but I checked the TU Dresden database and Christian's one isn't on there (I guess you probably already checked that too). The most likely explanation is that it has been embargoed - this is done sometimes when a dissertation includes IP owned or funded by a company (in this case Klippel). Embargoes are usually time-limited to 5 years or less, so it may become available at some point, but presently it appears that it still isn't. In the meantime, it appears that what you are doing is sensible (looking at the conditioning of the matrices, etc.)*Dr. Hargreaves, welcome to ASR!
I think/believe a significant chunk of the Klippel NFS development work was done by Christian Bellmann at TU Dresden (Dresden Technical University) and later at Klippel. The only mentioning to his thesis at TU Dresden I could find was from AudioXpress articles he coauthored (links below, see "About the Authors"). I, however, was not able to find a copy of his thesis online.
Practical Test & Measurement: The Proof Is in the Testing
The Klippel Near Field Scanner System (NFS) was used for the critical testing and measurement of Holoplot's X1 Matrix Array compensation algorithm capabilities. All our readers who have appreciated the story of the development of the unique Holoplot technology, published in the September 2022...audioxpress.com
Directivity Measurement of In-Wall Loudspeakers
In this article, the authors share their research on new requirements for the measurement of in-wall installation speakers. The directional characteristics are an especially important factor, because the direction in which a loudspeaker emits sound highly affects the interaction with the...
Bellmann is also one of the authors of this early Klippel NFS presentation.
My approach is slightly different in that the theory is based on pressure and particle velocity measurements on a common surface, as can be done with a Microflown probe. But much of the detail once one gets past the initial theory is much the same.
* though reading between the lines of the Klippel paper I linked to, it seems there they likely have some sort of optimisation algorithm running to decide where to adaptively add scan points once an initial scan is done, which I didn't see mention that you are currently doing and they don't say how they do. I should add that this is speculation though - I've seen a NFS, but haven't used the software or seen it running.
jahargreaves
New Member
- Joined
- Jun 2, 2026
- Messages
- 4
- Likes
- 17
Thanks. Yes this is much like what I've been doing too. We did more than is in the conference paper I shared - including some measurements - but to my discredit I've never managed to find time to publish this. Have you published any articles on your approach @d.fapinov?Hello @jahargreaves,
I think you'll find this interesting:
lah-scanner/documentation/centre_origin_help.md at main · dfapinov/lah-scanner
Python pipeline for 3D loudspeaker measurement and sound field reconstruction. Processes impulse responses on a spatial grid, performs spherical harmonic expansion (SHE), separates internal/externa...
No way, I am only an amateur.
jahargreaves
New Member
- Joined
- Jun 2, 2026
- Messages
- 4
- Likes
- 17
Well a very able amateur from the looks of it!
mwmkravchenko
Senior Member
Thanks for the papers and the link to the videos. More to learn from!
@NTK
We certainly did look! I believe the answer is that the thesis has been embargoed.
Mark
OP
- Thread Starter
- #592
We may not know all the secret sauces Klippel has cooked up, but with the excellent work by Tom, Jan, and Dimitri, I am confident that they'll get (or already are getting) something very close, if not equivalent, and definitely highly usable.
A lot of testing , like how near to the DUT, no of grid points, adjusting weight in point spread, DUT vertical and horizontal , the famous 'n' , and more parameters. I just did a reorganisation of the total HALS volume i have, 57GB, of which 54GB of measurement results.
And continuing with more measurements in the role of a user.
With one computer with stable sw for measuring different DUT's and one computer to test new software version when asked for by Tom or Dimitri.
We have daily feedback on progress , issues or lately also usability of the UI of the programs.
Luckily i am retired and above all my wife is very tolerant
And now that there are persons now building a scanner, we hope to get more feedback.
And continuing with more measurements in the role of a user.
With one computer with stable sw for measuring different DUT's and one computer to test new software version when asked for by Tom or Dimitri.
We have daily feedback on progress , issues or lately also usability of the UI of the programs.
Luckily i am retired and above all my wife is very tolerant
And now that there are persons now building a scanner, we hope to get more feedback.
@jahargreaves , i believe one of your papers is about measuring cardioid speakers.
In a week or 2 i will measure the Directiva R2 of @TimVG . Any advice to share?
In a week or 2 i will measure the Directiva R2 of @TimVG . Any advice to share?
Surprising many years before now I thought about studying acoustics at Salford when I was in UK. But, it was too academic for me and I wanted to do more things with my hands.
I did quick look at the paper, I should look more deepy. It is very interesting about the problem with finding acoustic source by minimizing di-pole term. I did have exactly this problem for bass reflex speakers that are a di-pole. The optimization wants to put one lobe of dipole at the origin so that it can be described by a monopole term. But then the other di-pole lobe requires higher orders. the problem is also like this for high frequency sources with complex lobe patterns and nearby diffraction.
I spent too much time trying to use different metrics to find the bets acoustic centre. It was one of the area I made many mistakes
Yes, fit error (residual) was the best.
I had an assumption that with both internal and external sources, if I use total fit error as metric to optimize that would not work to find only the internal source centre of expansion because it will be corrupted by fitting external sources. This was a wrong assumption, I think because good SNR with near-field measurement. But, at that time I was working only with synthetic data.
I tried to optimize mode sparsity, mode compactness, mean harmonic degree, entropy, matrix condition... I also found that ridge regularization was effective to get better results with lower order N testing of complex sources.
At the end I did get some very interesting images and learned about optimization types like differential evolution, nelder-mead and a vector search that I made to work with the concentric ring patterns in the data I had.
It also seemed to me that because the goal is good sound field separation in my case, simply optimizing low fit error is not enough - we can have low fit error but unstable matrix condition and poor field separation.
Here is some of the very nice image I got. It looks almost like diffraction x-ray I know without details, these image can not mean too much - I actually forgot a lot of the detail now, but they are still very interesting to me. We must remember these image do not show SPL or sound, it is plotting conditions of the SHE math
Below, real data genelec seaker.
Below 8.5KHz residual error synthetic source with directivity.
Below top down search view.
I did quick look at the paper, I should look more deepy. It is very interesting about the problem with finding acoustic source by minimizing di-pole term. I did have exactly this problem for bass reflex speakers that are a di-pole. The optimization wants to put one lobe of dipole at the origin so that it can be described by a monopole term. But then the other di-pole lobe requires higher orders. the problem is also like this for high frequency sources with complex lobe patterns and nearby diffraction.
I spent too much time trying to use different metrics to find the bets acoustic centre. It was one of the area I made many mistakes
Yes, fit error (residual) was the best.I had an assumption that with both internal and external sources, if I use total fit error as metric to optimize that would not work to find only the internal source centre of expansion because it will be corrupted by fitting external sources. This was a wrong assumption, I think because good SNR with near-field measurement. But, at that time I was working only with synthetic data.
I tried to optimize mode sparsity, mode compactness, mean harmonic degree, entropy, matrix condition... I also found that ridge regularization was effective to get better results with lower order N testing of complex sources.
At the end I did get some very interesting images and learned about optimization types like differential evolution, nelder-mead and a vector search that I made to work with the concentric ring patterns in the data I had.
It also seemed to me that because the goal is good sound field separation in my case, simply optimizing low fit error is not enough - we can have low fit error but unstable matrix condition and poor field separation.
Here is some of the very nice image I got. It looks almost like diffraction x-ray
I know without details, these image can not mean too much - I actually forgot a lot of the detail now, but they are still very interesting to me. We must remember these image do not show SPL or sound, it is plotting conditions of the SHE math Below, real data genelec seaker.
Below 8.5KHz residual error synthetic source with directivity.
Below top down search view.
Attachments
Last edited:
- Joined
- Sep 16, 2019
- Messages
- 1,418
- Likes
- 3,375
Looking forward to the results.
From my own measurements, you should see something going from wide cardioid to super cardioid - the effect is quite frequency dependent (as with all other "cardioid" systems).
I do not know if it is worth posting the same here and on DIY Audio.
Results for some experiments Tom suggested.
Doing multiple solve with different speed of sound constant.
That is similar as a temperature range of 20C. So it seems not too sensitive.
The lower points grid shows some wavelength related minima. That makes it difficult to optimize using only a small number of test frequency.
Also testing position sensitivity.
Fixed offset is not sensitive in Z and Phi axis - that is because acoustic origin optimizer can already compensate. Z is directly optimized. Phi can be optimized as mostly combination of X and Y shift. I will not waste space with plots.
That means R offset is the only one with sensitivity. This is like if the microphone is meant to be at 100mm but it is really 102mm. Such as if the diaphragm is not exactly at the mic tip
This shows the sound field separation ability at high frequency. In this test it was more clear than residual fit error.
R Offset
It looks like +5mm offset is a help.
Now random points jitter. This is different than fixed offset, it is more like if we have lost motion and backlash on the axis - some points are a little too froward, some are a little too back. Here the meaning of "1" means +/-1mm.
R Jitter
Z Jitter
Phi Jitter
This is degrees, not mm. So it does probably depend on the radius of measurement grid. Here the radius is small, 200mm. That means 1 degree arc length is about 3.5mm
In total I think it shows that position is tolerant to about 1-2mm total 'jitter'. 2mm can be seen having effect but it is not horrible.
Results for some experiments Tom suggested.
Doing multiple solve with different speed of sound constant.
That is similar as a temperature range of 20C. So it seems not too sensitive.
The lower points grid shows some wavelength related minima. That makes it difficult to optimize using only a small number of test frequency.
Also testing position sensitivity.
Fixed offset is not sensitive in Z and Phi axis - that is because acoustic origin optimizer can already compensate. Z is directly optimized. Phi can be optimized as mostly combination of X and Y shift. I will not waste space with plots.
That means R offset is the only one with sensitivity. This is like if the microphone is meant to be at 100mm but it is really 102mm. Such as if the diaphragm is not exactly at the mic tip
This shows the sound field separation ability at high frequency. In this test it was more clear than residual fit error.
R Offset
It looks like +5mm offset is a help.
Now random points jitter. This is different than fixed offset, it is more like if we have lost motion and backlash on the axis - some points are a little too froward, some are a little too back. Here the meaning of "1" means +/-1mm.
R Jitter
Z Jitter
Phi Jitter
This is degrees, not mm. So it does probably depend on the radius of measurement grid. Here the radius is small, 200mm. That means 1 degree arc length is about 3.5mm
In total I think it shows that position is tolerant to about 1-2mm total 'jitter'. 2mm can be seen having effect but it is not horrible.
I think it is if it's not too much trouble.
I've been thinking about the results.
The biggest key note is not all coordinate errors are the same sensitivity. Radial (R) errors have far the most negative effect on fit quality. Z errors are medium sensitive, and Phi errors are least sensitive.
R most sensitive > Z middle sensitive > Phi least sensitive
At first look the Z plot looks as bad as R, but I noticed I made the graph for Z until 10mm and the R and Phi only to 5mm / deg. Taking that, R is clearly the most sensitive coordinate.
We can think about why. Radius changes directly affects distance to the source, so it directly changes phase and field decay. It means we are measuring a different wavefront than intended. That's why the fit error with R grows with frequency. It is a direct radial error.
Phi errors are surprising insensitive. Even though a 1 degree movement is 3.5mm arc length in this test case (that is quite large if it is radial change) it may not be a big problem. Position error mostly moves microphone to a different position on the same wavefront rather than a different one (same expansion, different angle). The field seems much more tolerant of this error probably because the angular detail in the field is smooth. Maybe a highly directive source will be more sensitive but I think never so much as R.
It is good to think about how a internal source pattern expands – the field has a angular pattern, so even though 1 degree is a bigger arc length at large radius, it does not mean more error in the field sampling. The field also grows with distance and the detail spreads out. This means the long lever effect of Phi may not be so bad! Absolute error is magnified by a lever, but angle error is the same.
Z errors are mixture of radial and angular error. It part changes distance to source and part changes angle. So it is medium sensitive.
It is very convenient, I think. The axis that is most inaccurate in absolute position is less sensitive.
The biggest key note is not all coordinate errors are the same sensitivity. Radial (R) errors have far the most negative effect on fit quality. Z errors are medium sensitive, and Phi errors are least sensitive.
R most sensitive > Z middle sensitive > Phi least sensitive
At first look the Z plot looks as bad as R, but I noticed I made the graph for Z until 10mm and the R and Phi only to 5mm / deg. Taking that, R is clearly the most sensitive coordinate.
We can think about why. Radius changes directly affects distance to the source, so it directly changes phase and field decay. It means we are measuring a different wavefront than intended. That's why the fit error with R grows with frequency. It is a direct radial error.
Phi errors are surprising insensitive. Even though a 1 degree movement is 3.5mm arc length in this test case (that is quite large if it is radial change) it may not be a big problem. Position error mostly moves microphone to a different position on the same wavefront rather than a different one (same expansion, different angle). The field seems much more tolerant of this error probably because the angular detail in the field is smooth. Maybe a highly directive source will be more sensitive but I think never so much as R.
It is good to think about how a internal source pattern expands – the field has a angular pattern, so even though 1 degree is a bigger arc length at large radius, it does not mean more error in the field sampling. The field also grows with distance and the detail spreads out. This means the long lever effect of Phi may not be so bad! Absolute error is magnified by a lever, but angle error is the same.
Z errors are mixture of radial and angular error. It part changes distance to source and part changes angle. So it is medium sensitive.
It is very convenient, I think. The axis that is most inaccurate in absolute position is less sensitive.
