Neumann KH 80 DSP Speaker Measurements: Take Two

[Pio2001]

I verified this in the last few hours.

Thank you Amir !

To understand all this, reading this presentation is very useful : https://www.klippel.de/fileadmin/us..._3D_Sound_Field_using_Near_Field_Scanning.pdf

As I understand it, sound field separation and sound field expansion are two different things.
Looking at pages 33, 37 and 39 in the above document, it seems to me that sound field separation is used at low frequencies (page 37), but that sound field expansion is used at all frequencies (page 33, the order of expansion has a strong effect at 10 kHz, and page 39, it says "wave expansion interpolates..." about data at 2.5, 5, 8 and 10 kHz ).

It makes sense: the scan is made at very short distances (30 cm from the speaker). Sound field expansion is useful to predict the response far from the speaker, even at high frequencies, especially off axis.

Thus, if I understand correctly, the system does, at low frequences:
Measurements -> sound field separation -> sound field expansion
And at high frequencies:
Measurements -> gating -> sound field expansion

In order to understand the fitting error graphs, we must look at page 32 and 41 (funny : they too measured a Neumann KH-80 for their presentation )
The fitting error graphs are generated after each measurement using redundant data. They show how large is the variation coming from redundant data taken during the measurement.

In order to know the effet on the resulting frequency response, we must convert the error level in dB (from the fitting error graph) into an absolute number (10^(x/20)), then add 1 and divide by 1 in order to get the error ratio instead of error level, then convert back into dB:
Error on the FR graph (dB) = 20 log (1+10^(x/20)) where x is the error level on the fitting error graph.

It gives
Fitting error -20 dB -> FR curve accuracy 0.8 dB
Fitting error -10 dB -> FR curve accuracy 2.4 dB
Fitting error -5 dB -> FR curve accuracy 3.9 dB
Fitting error -4 dB -> FR curve accuracy 4.2 dB
Fitting error -2.5 dB -> FR curve accuracy 4.9 dB

Now, we can have a look at the measurements:

The default 500+ point and order 10 gets accuracy to about 6 kHz. Above that error climbs rapidly due to complexity of the soundfield:

View attachment 49245

Given the same number of points, the highest order I can increase in software is 14 and this is the results:

View attachment 49246

The error is pushed down in higher frequencies. Results are still not very good above 9 kHz.

Converting error level into accuracy of the frequency response, We can see that with order 10, the accuracy of the FR is
0.9 dB at 6000 Hz
1.6 dB at 7000 Hz
3.2 dB at 8000 Hz
3.9 dB at 9000 Hz
4.2 dB at 10 000 Hz
4.9 dB at 14 000 Hz

With order 14 the accuracy becomes
0.3 dB at 6000 Hz
0.4 dB at 7000 Hz
0.7 dB at 8000 Hz
1.4 dB at 9000 Hz
2.4 dB at 10 000 Hz
3.9 dB at 14 000 Hz

Comparing your two frequency response graphs, at orders 10 and 14, we can see that they are identical within 1 dB below 8000 Hz, they differ of 1 dB at 10 Khz, and there is a 2 dB difference at 14 kHz.
The consistency is much better than what is predicted by the fitting error. Are the fitting error graphs taken from the tweeter's measurement or from the center point measurement ? The second case would make more sense, but maybe the fitting error is a conservative number, and the averarge value is still accurate in spite of possible fitting errors.

Don't panic! The data I show you is independent of this:


View attachment 49247

Gating is used above 2 kHz where we have plenty of error margin in the previous graph.

I don't think so. As seen above, the absence of sound field separation (gating) doesn't mean the absence of sound field expansion (with fitting error).

Indeed, I compared the generated frequency response graphs and they are essentially identical regardless of expansion order (these are for the on-tweeter-axis):

View attachment 49248
View attachment 49249

That's right. But according to what you posted previously, high order expansion becomes critical when the tweeter point is not properly defined. Here, with the tweeter point properly set, high order expansion is not necessary.

The other graphs, with the wrong tweeter point, however, might look very different at order 10 and order 14.

I will resort to that if fitting error drops too low in the ungated area on complex speakers but for now, I think we are good with the number of measurement points.

In spite of the large fitting error (if it is indeed the fitting error from the measurements taken with the proper tweeter point) it seems that the curves are quite stable indeed.
Especially if we keep in mind that the green dashed curve (listening window) is probably more important than the black solid one (on-axis). That one (listening window, green) is identical up to 14 kHz, and it varies 1 dB at 20 kHz between order 10 and order 14.

 

[Pio2001]

Okay here we go again thanks to @Pio2001 pointed out the simple problem into post 146 so even myself got it

View attachment 49268

Thanks for the new graph, Byrtt !

 

[digitalfrost]

To understand all this, reading this presentation is very useful : https://www.klippel.de/fileadmin/us..._3D_Sound_Field_using_Near_Field_Scanning.pdf

From the picture on page 28, it seems to me Klippel themselves didn't measure the KH80 with the mic at tweeter height, but at the specified acoustic center. Am I missing something?

 

[thewas]

From the picture on page 28, it seems to me Klippel themselves didn't measure the KH80 with the mic at tweeter height, but at the specified acoustic center. Am I missing something?

A snapshot must not necessarily show the reference axis as measurements are taken all around the speaker, also on that slide they are showing a total sound power measurement where the reference axis doesn't matter as its equally integrated all over the surface.
PS: By the way I think that is a KH120 and not KH80.

 

[Pio2001]

PS: By the way I think that is a KH120 and not KH80.

oooh, you're right. The tweeter's size looks different compared to the speaker's size.

 

[ctrl]

It gives
Fitting error -20 dB -> FR curve accuracy 0.8 dB
Fitting error -10 dB -> FR curve accuracy 2.4 dB
Fitting error -5 dB -> FR curve accuracy 3.9 dB
Fitting error -4 dB -> FR curve accuracy 4.2 dB
Fitting error -2.5 dB -> FR curve accuracy 4.9 dB

Thanks Pio2001!

This seems to be the missing piece of the puzzle to explain the disturbances (ripple) in the range of 5kHz to 15kHz that can be detected in more or less all measurements.

Even at 5kHz the fitting error alone is 0.8dB according to your calculations and increases rapidly as Amir has also confirmed - besides the usual errors of mic, mic preamp and audio interface (if not measured in dual channel mode).

This also explains why all normalized measurements look smoother, since the ripple error changes only slightly at different angles and thus is similarly "strong" in each measurement, this error is eliminated during normalization.

So you can say that the Klippel NFS measurements in the low frequency range have advantages over ancheoic chamber measurements.

But in the range above 5kHz the measurements are much less accurate than measurements in ancheoic chambers, probably even less accurate than well done gated DIY measurements?

 

[napilopez]

Just wanted to chime in and say I'm glad this thread and the overall Neumann discussion has evolved from a heated debate about measurement procedures and results to a constructive one about understanding how the NFS works. I think we're all smarter and more educated about the NFS measurements for it!

 

[amirm]

I don't think so. As seen above, the absence of sound field separation (gating) doesn't mean the absence of sound field expansion (with fitting error).

Yeh after I post that I did some other analysis and it seems that expansion is used for full response. There is good news however in that the computation of fitting error seems very conservative. The redundant points must have been picked to show some kind of worst case as the impact is quite minimal on frequency response measurements.

To wit, let's look at the Harbeth 30 original measurement error:

This says in the 2 to 3 KHz we exceeded the -20 dB criteria and at 20 kHz, lost it good bit more. Here is the frequency response as reported originally:

I just recomputed this using higher order expansion resulting in this fitting error plot:

We gained quite a bit of margin and we are essentially below -20 dB for the entire graph. Here is the new spin data:

At 3 kHz, that dip is a bit better. At 20 kHz, the peak is a bit lower. But overall picture of the two graphs is essentially identical.

I am going to re-test the KH80 to get the low fitting error to see what impact it has there.

 

[MZKM]

Yeh after I post that I did some other analysis and it seems that expansion is used for full response. There is good news however in that the computation of fitting error seems very conservative. The redundant points must have been picked to show some kind of worst case as the impact is quite minimal on frequency response measurements.

To wit, let's look at the Harbeth 30 original measurement error:

View attachment 49364

This says in the 2 to 3 KHz we exceeded the -20 dB criteria and at 20 kHz, lost it good bit more. Here is the frequency response as reported originally:

View attachment 49365

I just recomputed this using higher order expansion resulting in this fitting error plot:

View attachment 49366

We gained quite a bit of margin and we are essentially below -20 dB for the entire graph. Here is the new spin data:

View attachment 49367

At 3 kHz, that dip is a bit better. At 20 kHz, the peak is a bit lower. But overall picture of the two graphs is essentially identical.

I am going to re-test the KH80 to get the low fitting error to see what impact it has there.

For either the Harbeth or the Neumann once you re-test it with the higher order, can you share the spin data so that I can see if the preference score is altered in any meaningful way?

 

[amirm]

For either the Harbeth or the Neumann once you re-test it with the higher order, can you share the spin data so that I can see if the preference score is altered in any meaningful way?

The Harbeth is returned already but fortunately just changing the numerical analysis the error drops below -20 dB (1%). Here it is:

The Neumann is measuring with 1000 points which takes 2 hours. I will have to post process and then will post it.

 

[MZKM]

The Harbeth is returned already but fortunately just changing the numerical analysis the error drops below -20 dB (1%). Here it is:

The Neumann is measuring with 1000 points which takes 2 hours. I will have to post process and then will post it.

Going to see Birds of Prey now, will post data in a few years.

 

[edechamps]

Thank you, @amirm, for indulging us with this second set of measurements. Looks like your curious scientist side prevailed over your frantic reviewer side To me that's a good thing - we are learning a ton of stuff in this thread that will be useful for future reviews.

The second set of measurements are somewhat of a plot twist. They definitely compel more questions than they answer. As a reminder, the initial hypotheses were that (1) the bass dip was caused by excessive measurement level and (2) the upwards treble tilt was caused by use of the wrong reference axis.

Regarding the bass dip (1), clearly the hypothesis is wrong. Something else is causing the discrepancies between the NFS measurement and third party anechoic measurements, and we don't know what it is. One could argue that the NFS is supposed to be better at measuring bass frequencies because anechoic chambers are not truly anechoic in the lower parts of the spectrum. That might well be true, but that would mean that all the anechoic chambers that were used in third party measurements are deviating in the exact same way, which seems implausible to me.

To be fair I'm not that worried about the bass issue when it comes to future reviews, because the difference is small and people who care about such small differences will likely use room EQ which will implicitly compensate for them anyway. I am, however, slightly worried that the measurements appear to deviate in ways that no-one seems to be able to explain. I do believe it would still be useful for @amirm to report it to Neumann and Klippel. If the NFS is right, Neumann might be interested to know that a speaker they think measures flat is not actually flat. If the anechoic measurements are right, Klippel might be interested to know that the NFS seems to have an accuracy problem in this case.

Regarding the treble tilt (2), let me try to summarize the theory that people in this thread seem to be converging on: the NFS doesn't like it when we force it to measure from a point that is not the highest frequency emitter in the system (because of mathematical considerations that go way above my head). In fact, it hates it so much that, with the number of measurement points Amir is using, accuracy goes down the drain above 7 kHz or so. The proposed solution is to keep measuring based on the location of the tweeter, but then adjust the post-processing done in software such that the graphs show the behaviour of the appropriate reference axis (which might or might not be the tweeter). If I understand correctly this is what @amirm set out to do when he said:

You can set this mechanically or in the software. The latter is a pain as you have to compute the coordinates to put in there. I will make another measurement by setting this using the robotic system and see what difference it makes.

Unless I missed it Amir did not post the results of that experiment yet. Even if it works, this begs the question of what to do if we come across an "exotic" speaker that doesn't have a well-defined, single, forward-firing tweeter. I guess in this specific case we can just "brute force" it by cranking the number of measurement points way up?

 

[amirm]

Going to see Birds of Prey now, will post data in a few years.

Years? I thought I was bad with my review pace. Enjoy the movie.

 

[amirm]

Unless I missed it Amir did not post the results of that experiment yet. Even if it works, this begs the question of what to do if we come across an "exotic" speaker that doesn't have a well-defined, single, forward-firing tweeter. I guess in this specific case we can just "brute force" it by cranking the number of measurement points way up?

That's correct. That experiment is in play as I measure at 1000 points with KH80. I am hoping this is as high as we need to go. But if not, then the number can be increase again at the high cost of measurement time. Then again it is automated so it is not too bad at the end of the day.

 

[LeftCoastTim]

we are learning a ton of stuff in this thread that will be useful for future reviews.

I completely agree. Neumann claims extremely tight control in frequency response as well as unit-to-unit variation. By measuring two samples, we are verifying their claims as well as learning more about the NFS.

KH80 #2 is my unit, and I'm flattered it's getting a 3rd scan by the NFS

 

[edechamps]

Oh, and by the way, two questions came to mind while reading this thread:

The first one, and I'm sorry as maybe this was answered already, is: when @amirm posts something like a Spinorama, is the graph showing estimated far field conditions, as in, infinite distance from the speaker, and then normalizing for SPL at a given distance? Or is it actually showing the expected response at some fixed, reasonable distance from the speaker? If it's the latter, what is the distance used?

The second one is: how does the NFS deal with the fact that the microphone is not perfectly omnidirectional? If the NFS positions the microphone such that high frequency sound hits the microphone at some wide angle, the microphone will not measure the true sound pressure at that point. Will that error show up in the resulting data, or does the NFS somehow compensate for that? Can this explain some of the phenomena we're observing, given that, IIRC, @amirm said that the microphone sometimes gets very close to the speaker, which would de facto increase angles of incidence on the microphone capsule?

 

[amirm]

One thing we need to find out about the anechoic measurements of KH80 is what smoothing/frequency resolution they are using. I have not seen that document. The S&R magazine just says what the sample rate of audio was from the microphone, not the resulting resolution of their graph. Has anyone seen otherwise?

 

[amirm]

The first one, and I'm sorry as maybe this was answered already, is: when @amirm posts something like a Spinorama, is the graph showing estimated far field conditions, as in, infinite distance from the speaker, and then normalizing for SPL at a given distance? Or is it actually showing the expected response at some fixed, reasonable distance from the speaker? If it's the latter, what is the distance used?

Far field is defined by simple 6 dB SPL reduction when doubling the distance. So there is no change in response other than the vertical scale changes.

This *is* an issue for anechoic chamber because they are not able to often get far enough from a speaker to be in true far-field.

Currently the vertical SPL absolute value for powered speakers is wrong. I know why it is computed wrong but don't know the reasoning. The relative levels (i.e. non-flatness) in response is correct however.

 

[amirm]

The second one is: how does the NFS deal with the fact that the microphone is not perfectly omnidirectional? If the NFS positions the microphone such that high frequency sound hits the microphone at some wide angle, the microphone will not measure the true sound pressure at that point. Will that error show up in the resulting data, or does the NFS somehow compensate for that? Can this explain some of the phenomena we're observing, given that, IIRC, @amirm said that the microphone sometimes gets very close to the speaker, which would de facto increase angles of incidence on the microphone capsule?

It has no way of knowing microphone characteristics. It however checks its own work with respect to field expansion as I have shown.

In an anechoic chamber, you have a similar problem as the speaker is rotated and angle of incident changes relative to microphone. I don't know anyone who is correcting for that. Harman uses microphone array and there, each mic is likely different from another by a bit as well. Here is a picture I took while I was there:

As I keep saying, we need to be focused on the high level picture. Slight differences are simply not material with respect to predicting listening response in their own room.

 

[JIW]

One thing we need to find out about the anechoic measurements of KH80 is what smoothing/frequency resolution they are using. I have not seen that document. The S&R magazine just says what the sample rate of audio was from the microphone, not the resulting resolution of their graph. Has anyone seen otherwise?

The measurements for S&R are done by Anselm Goertz who states the resolution as 1 Hz or lower.

Source (in german). Translated to english.