The NFS uses a time window above ~1kHz to remove room reflections, but below that, it takes removing room reflections a step further: by measuring on a pair of concentric cylindrical "surfaces" around the speaker, it knows what sound is going out through the measurement surfaces (from the speaker) and what sound is coming in through those surfaces (from the room) and is then able to remove the sound from the room out of the measurement. As you know, a short time-gate reduces the resolution of your measurements, and even outside 6' in the air, detail below ~1kHz is seriously impaired.
And yes, you are correct that the robot part of it also gets you the 3D balloon plots.
Thanks!
A friend sent me there. And I have read the thread until the wee hours of the morning. Even added a little 2 cents in today.
But can also it do a full range measurement through the means it like does for below 1khz? I mean it would save you the lost resolution above 1khz associated with gating...
As the math behind gating works out, the resolution loss decreases the higher in frequency you go. So even with a 2 or 3 millisecond window by the time you get up to about 1.5kHz-2kHz the differences between "true" anechoic and gated measurements range from academic to nonexistent.
But as far as using sound field separation at high frequency goes, the quote from the movie Hunt For Red October is applicable: "possible, but not recommended." The math behind the process can technically do it, but at a higher computational load and much higher likelihood of errors. Hence why Klippel's system has a hard limit to how high in frequency it will perform sound field separation.
aslepekis
I think that you have this slightly wrong. Sound Field Separation (modal analysis) is done at all frequencies, but cancelling reflections by using dual spherical measurements is limited to LFs, below the frequency that gating becomes erroneous. It is not required when gating is used.
I think that you have this slightly wrong. Sound Field Separation (modal analysis) is done at all frequencies, but cancelling reflections by using dual spherical measurements is limited to LFs, below the frequency that gating becomes erroneous. It is not required when gating is used.
Oops, my mistake. Thanks for the clarification! I had been under the impression that Sound Field Separation was the term for the process of canceling room reflections and Multi-Pole Expansion was the term for the modal analysis that is performed at all frequencies. But I see on Klippel's website they refer to the reflection cancelation as Direct Sound Separation.
I thought I'd share an update to this project:
With thanks to a suggestion by user NTK over at Audio Science Review, I tried an experiment tonight (the technique is detailed here and here, and conceptually very similar to this). The method is to sum multiple impulse responses from along an axis radiating from the speaker under test, the idea is that randomly differing reflections will be smoothed out but the desired response from the loudspeaker will remain. However, previous examples of this technique have been performed in rooms, so a minimum of six reflective surfaces and modal behavior of the room are needing to be overcome. This experiment was to see how the technique would fair outside with only a reflection from the ground to contend with.
It was a quick and dirty experiment in my back yard, so road construction, neighborhood and bird noises are in there. Plus I didn't want to run the measurement sweeps too loud since it was after 8:00PM... and I wasn't super careful with mic placement. So this ain't laboratory grade. Nevertheless, I think the results are still significant. The process was to sum 40 individually taken IR's in Audacity and export that sum as a .WAV file for import into REW. The measurements themselves were taken on roughly the tweeter axis of a Behringer B2030P starting at 37" away from the speaker and ending at 71" away. The speaker and microphone were about 7' above the ground. Below are the results with no smoothing or IR windowing. The black line is the processed data; the gray line is a single one of those measurements used in the process for comparison; the red line is a Behringer B2030P measured by Amirm at Audio Science Review on his Klippel NFS; the green line is another one of my 2030P's measured several years ago in my living room just to emphasize the difference. Again, no IR window or smoothing has been applied to any of this data.
As another comparison, in the linked paper by Daniel Krol, where this experiment was inspired, 100 measurement points and algorithmic processing were used to get the following in comparison between anechoic and this quasi anechoic in room measurement method:
More investigation needs to be made to refine the process, and to see just how few data points are needed to get useful data, but it is a promising technique for DIY'ers to get a better look at what a speaker is doing below 1kHz... even if it's not a Near Field Scanner on a shoestring.
With thanks to a suggestion by user NTK over at Audio Science Review, I tried an experiment tonight (the technique is detailed here and here, and conceptually very similar to this). The method is to sum multiple impulse responses from along an axis radiating from the speaker under test, the idea is that randomly differing reflections will be smoothed out but the desired response from the loudspeaker will remain. However, previous examples of this technique have been performed in rooms, so a minimum of six reflective surfaces and modal behavior of the room are needing to be overcome. This experiment was to see how the technique would fair outside with only a reflection from the ground to contend with.
It was a quick and dirty experiment in my back yard, so road construction, neighborhood and bird noises are in there. Plus I didn't want to run the measurement sweeps too loud since it was after 8:00PM... and I wasn't super careful with mic placement. So this ain't laboratory grade. Nevertheless, I think the results are still significant. The process was to sum 40 individually taken IR's in Audacity and export that sum as a .WAV file for import into REW. The measurements themselves were taken on roughly the tweeter axis of a Behringer B2030P starting at 37" away from the speaker and ending at 71" away. The speaker and microphone were about 7' above the ground. Below are the results with no smoothing or IR windowing. The black line is the processed data; the gray line is a single one of those measurements used in the process for comparison; the red line is a Behringer B2030P measured by Amirm at Audio Science Review on his Klippel NFS; the green line is another one of my 2030P's measured several years ago in my living room just to emphasize the difference. Again, no IR window or smoothing has been applied to any of this data.
As another comparison, in the linked paper by Daniel Krol, where this experiment was inspired, 100 measurement points and algorithmic processing were used to get the following in comparison between anechoic and this quasi anechoic in room measurement method:
More investigation needs to be made to refine the process, and to see just how few data points are needed to get useful data, but it is a promising technique for DIY'ers to get a better look at what a speaker is doing below 1kHz... even if it's not a Near Field Scanner on a shoestring.
This is very interesting. Was this private correspondence or am I able to read this on the forum? I have a large yard an I live in the country so doing this is not to difficult. Right now I set up on a 14 foot high ladder for anechoic to about 55 hertz. But it is a bit of a pain to do this. Your setup is much easier.
It's all in the public forum (here and here), but those linked papers by Daniel Krol have more detailed information. I believe they are public access, but feel free to PM me if they are not. I did all the measurements in REW without a timing reference since it will then auto-center the measurement IR peak at t=0 and that makes the whole process of summing all the IR's easier, but more effort would be needed to align the IR's properly in the summing process when a timing reference is used.
Since you have a far better environment to try this in, I'm very interested to see what results you can get with this setup if you do try it out.
It's noisy in my neighborhood today, but when I get another chance to experiment with this, I'm curious to find out how far from the ground the speaker actually needs to be since it's a lot easier to align everything if the speaker only needs to be elevated, say, three feet rather than seven.
I'm glad you found it interesting!
Since you have a far better environment to try this in, I'm very interested to see what results you can get with this setup if you do try it out.
It's noisy in my neighborhood today, but when I get another chance to experiment with this, I'm curious to find out how far from the ground the speaker actually needs to be since it's a lot easier to align everything if the speaker only needs to be elevated, say, three feet rather than seven.
I'm glad you found it interesting!
Another update:
I used the data I collected last night to try and get an idea of how many measurement points are needed to get acceptable results. Again, the measurement points were taken at approximately 0.8" intervals between 37" and 71" away from a Behringer B2030P, but this time only selected IR's were summed together.
In descending order, black is all 40 points; gray is 20 points (just the odd numbered IR's); dark blue is 8 points (just the numbered IR's that end in 5); light blue is 4 points (just the numbered IR's that end in a 0); dark red is two points (just IR #1 and #40); light red is one point (IR #1).
I am very surprised just how passable even just 4 measurement points is.
Note that for this method to lessen the effects of the ground reflection, distance of microphone spacing from the speaker is imperative. So taking 60 IR's over a span of 3" won't get you anywhere (at least not below 1-2kHz).
I used the data I collected last night to try and get an idea of how many measurement points are needed to get acceptable results. Again, the measurement points were taken at approximately 0.8" intervals between 37" and 71" away from a Behringer B2030P, but this time only selected IR's were summed together.
In descending order, black is all 40 points; gray is 20 points (just the odd numbered IR's); dark blue is 8 points (just the numbered IR's that end in 5); light blue is 4 points (just the numbered IR's that end in a 0); dark red is two points (just IR #1 and #40); light red is one point (IR #1).
I am very surprised just how passable even just 4 measurement points is.
Note that for this method to lessen the effects of the ground reflection, distance of microphone spacing from the speaker is imperative. So taking 60 IR's over a span of 3" won't get you anywhere (at least not below 1-2kHz).
This is an interesting idea.
I would think that one would still want to gate at HFs since it's so reliable and easy, but getting LF resolution has always been an issue. This technique may well be an answer. Given that then wouldn't a few point widely spaced be the best approach? Blended into a gated HF response.
How do you adjust for the level differences in the impulses? Just normalize every IR to have a peak value of 1.0?
I would think that one would still want to gate at HFs since it's so reliable and easy, but getting LF resolution has always been an issue. This technique may well be an answer. Given that then wouldn't a few point widely spaced be the best approach? Blended into a gated HF response.
How do you adjust for the level differences in the impulses? Just normalize every IR to have a peak value of 1.0?
I'm inclined to agree, and that would help address higher frequency ambient noise from leaves and birds or other things. High frequency data could even be obtained from indoor measurements if one wanted to. Given the performance of 8 points spaced over 34", I'd like to see what a few points spaced over 50" or more would do.
I didn't, actually. I had been thinking of normalizing the IR's like that when I exported them from REW, but for my initial experiment I just exported them as is. There was only about 4dB of level difference between the closest and farthest measurements.
I had a similar idea of doing that years ago.
Just doing it by hand was extremely tedious at best (and my programming skills aren't great unfortunately)
But nice to know they use it in their system as well.
I only don't really agree with the idea that your whole measurement will be garbage by bigger time windows.
(you said short, but I think you mean bigger time-gate)
Just by the fact that one does more than just one measurement.
So when looking at the nearfield response, burst decay envelope, (near field) distortion curve as well as the impedance and you get pretty much all the information you need to know, including all issues that will be at play.
I don't really see how that is very useful to be able to plot in the far-field (well 1 meter or so) frequency response?
Mostly because that set of measurements is only really important for the overall picture as well as for the directivity.
With the stitching method (and/or used with grounplane measurements) the results are extremely good and accurate.
But in the end I guess it's more about knowing information for designing a speaker vs reviewing speakers.
Anyway, what would already help is using cardiod microphones as well.
I think there is no reason at all to use omnidirection mics for this purpose to be honest, since we only have to measure the frequency response on one plane. It already gets rid of a lot of reflections from the back (wall) for example.
However, it's quite difficult to get a decent cardiod mic capsule.
Which has more to do with a chicken-egg story tbh.
Another idea is to actually mount microphones on the locations were the wave reflects from.
In that case there is some cross-correlation that we can do and some additional mathematical trickery.
With a good DSP can be done live or just in post later on.
Measuring with anti-parallel surfaces, also with respect to the source and microphone, will improve a lot as well.
Last edited:
Well I don't know but if we are using advanced computing anyway, it's not so difficult anymore to detect the best gate-time window? In case of Klippel, they already have minimum requirements for the size of the room if I am not mistaken?
So at that point you already know what to expect.
With a Hann 25% window (I believe that's the same as a Tukey 0.50 ?) you can easily get your lowest point around 170-200Hz. The very last bit won't have as much details, but at that point with stitching method is already taking over.
I can do a twin T filtered stepped sine wave of n number of points. It is tedious and terrible to listen to. But incredibly good at being immune to all other sound sources. The trick is to not choose to many points.
You can do ground plane measurements up to about 13khertz and reliably match an anechoic measurement. So for the bulk of what we need to measure if you have the conditions and the weather you are fine. I live in Canada. So I do this in the summer. Now up until about September, sometimes October. But you can't really do polars this way. At least I have never compared them to polars on my tall step ladder. Something I should do.