Another idea I had in mind, is to map just only the reflections.
Which might be possible to measure a driver as a dipole.
Here the cone/back of the driver, faces the floor/ceiling.
The mic will barely record any direct sound (since you're in the null of the dipole) but will record floor reflections.
(well that's the theory lol)
The bigger the distance, the better this works.
Well belatedly I have a few measurements that I can show to compare a multiple in room measurement via REW and a proper ground plane measurement. Please note that there will be an absence of floor and ceiling bounce effects on the groundplane just as you will be lissing this in a true anechoic chamber measurement.
So to set the measurement up.
This measurement room is basically 8 feet wide ( 2.5 metres ) 8 feet high and 16 feet long. The loudspeaker is placed at the least modal point in the room. The loudspeaker has a rear firing Passive radiator which was close miced as was the over all response in the front side and this was summed using the math functions available in REW. 25 measurement in 100 mm increments were made from 20mm distance on out. Then I applied a few of the functions available in REW to remove the reflections and sum nearfield and far field measurements.
Yes, in this room at this position this is what I measured. It should dive at 45 hertz. Wait for the groundplane measurements.
Ok now outside for recess.
Due to asymmetric tweeter location I could get a slight difference between a vertical orientation and a horizontal orientation. But it was not night and day different. Mic position made more of a difference actually.
Equivalent of the inside 0.5 metre measurement and I also did the correct 2 metre measurement to simulate a 1 metre 2Pi measurement.
So, what did this get me? Smoothed versus Groundplane measurement?
Have at it men!
So to set the measurement up.
This measurement room is basically 8 feet wide ( 2.5 metres ) 8 feet high and 16 feet long. The loudspeaker is placed at the least modal point in the room. The loudspeaker has a rear firing Passive radiator which was close miced as was the over all response in the front side and this was summed using the math functions available in REW. 25 measurement in 100 mm increments were made from 20mm distance on out. Then I applied a few of the functions available in REW to remove the reflections and sum nearfield and far field measurements.
Yes, in this room at this position this is what I measured. It should dive at 45 hertz. Wait for the groundplane measurements.
Ok now outside for recess.
Due to asymmetric tweeter location I could get a slight difference between a vertical orientation and a horizontal orientation. But it was not night and day different. Mic position made more of a difference actually.
Equivalent of the inside 0.5 metre measurement and I also did the correct 2 metre measurement to simulate a 1 metre 2Pi measurement.
So, what did this get me? Smoothed versus Groundplane measurement?
Have at it men!
Attachments
Thanks for sharing @Kravchenko_Audio!
Color me impressed that this "beamforming" method can be used indoors so effectively (that wart between 45Hz and 65Hz is VituixCAD's SPL trace not cooperating). One question on your indoor DSB measurement: is it RMS Averaged or Vector Averaged? I've been Vector Averaging mine and have noticed that it makes a difference at lower frequency.
Also, I like your Sienna, so much room to haul stuff!
I put 250,000 miles on mine before finally giving it away.
Color me impressed that this "beamforming" method can be used indoors so effectively (that wart between 45Hz and 65Hz is VituixCAD's SPL trace not cooperating). One question on your indoor DSB measurement: is it RMS Averaged or Vector Averaged? I've been Vector Averaging mine and have noticed that it makes a difference at lower frequency.
Also, I like your Sienna, so much room to haul stuff!
I put 250,000 miles on mine before finally giving it away.
I never mentioned that all the levels were taken at 85 db at the stated distance. So the last graph was 85db groundplane 2 metres. Not normalized from a nearfield measurement.
The measurements were normalized to a common point and them RMS I think. I tried a few different ones. And this was done on the 30 and 31st of August. What I have come to appreciate is that you can get pretty accurate measurements with REW. I never really tried it before. That and the UMIK are a great mix for serious measurement.
There is a lot I am still learning with the software. Old dog needs to learn more tricks I think. Woof er.....
The measurements were normalized to a common point and them RMS I think. I tried a few different ones. And this was done on the 30 and 31st of August. What I have come to appreciate is that you can get pretty accurate measurements with REW. I never really tried it before. That and the UMIK are a great mix for serious measurement.
There is a lot I am still learning with the software. Old dog needs to learn more tricks I think. Woof er.....
I'm late to the party and am doing a lot of reading up. This old quote got me thinking:
To my untrained eye it seemed that higher frequencies benefit from smaller gates, whereas low frequencies need a large gate to be shown at all in REW.
What if we make this gating adaptive? Small where it can be, large where is has to be? I made a proof-of-principle function in Octave and that resulted in this:
Comparison ungated REW vs adaptive gating:
Adaptive gating:
ps. The legend says seconds; that should of course be milliseconds.
To my untrained eye it seemed that higher frequencies benefit from smaller gates, whereas low frequencies need a large gate to be shown at all in REW.
What if we make this gating adaptive? Small where it can be, large where is has to be? I made a proof-of-principle function in Octave and that resulted in this:
Comparison ungated REW vs adaptive gating:
Adaptive gating:
ps. The legend says seconds; that should of course be milliseconds.
You may be late, but you are still welcome!
Your experiments are making me wish I could do more than simple arithmetic in Octave.
It is true that lower frequencies need a larger gate to provide acceptable resolution, however, in a reflective (i.e. a room) or hemi-anechoic (i.e. outside) environment, the reflections that need to be gated out of the measurement are full bandwidth and arrive to the microphone in the single-digit millisecond range. In other words, any reflection that is corrupting the high frequency will be doing the same to the low.
That said, you do appear to have less "grass" in your adaptive gating results, so it may be a useful method to apply in concert with the other methods discussed in this thread.
As mentioned by others above, this has been done before (HolmImpulse, JustMLS). I also imlemented this in MATAA a long while ago, but it felt like cheating. The thing is that at high frequencies, you're looking at the anechoic response of the speaker, whereas at low frequencies you're looking at the echoic response. The lower you go, the more the response will be determined by acoustics of the room, not those of the speaker.
Since we tend to be concerned with testing speakers (not rooms), we need to find ways to avoid the effects of the room, or at least to filter them out from the measurement data. This is what the "Klippel method" does.
I was thinking of about cheaper method that might work based on a conventional impulse response measurement from a single microphone. I am not saying that it would be as good as the "Klippel method", but it might at least be better than nothing. The idea is to split the impulse response in the anechoic and echoic part. The anechoic part would then be used to estimate its echo(es) present in the echoic part by using some sort of a regression model (not sure how exactly, but this might require some educated assumptions and heuristic approaches). This (estimated!) echo could then be subtracted from the echoic measurement to reconstruct (or guesstimate...) the anechoic residual in the echoic part of the measurement. Yes, that's just a wild idea, and I haven't really developed this, but I thought I'd throw it out there... Also, I wouldn't be surprised if this concept was already used in the past, either in acoustics or maybe also in other disciplines like seismics etc.
The best technique (other than the Klippel approach, which is really the Weingard approach) is to use linear regression on the impulse response to extend it through the reflection zone. I've tried this in many different ways and got nowhere. But it was used in a software package (name eludes me, Waslo, I think) did some software many years ago for measuring speakers.
What exactly do you mean by this? The term "linear regression" is used in very different ways, and without context it's sometimes a bit difficult to make sense of it.
Did you mean...
... to use the last anechoic sample and an imaginary sample with value 0 somewhere a t --> infinity, and then do a linear interpolation ("straight line") between these two samples?
... or something more complicated like a linear model that estimates the echo from the anechoic part? The trick will be be to figure out a good model.
Anyone checked out stuff for internet calls? At least Teams is very good suppressing any sounds other than speech, echo and all. I suppose there would be lots of stuff available, from python libraries to stuff I can't imagine. Or is it something completely different to what happens on low frequencies? Just highpass?
I have thought of that, although I haven't checked it out too deeply. But if inserting a Teams meeting into the measurement chain is what it takes, I'm all for it!
If (a big if) I understand member @mikets42 comments on echo cancelation (in papers linked in his thread here), the non-LTI components produced by the drivers will work against you. Also, if anyone hasn't checked out that thread, I recommend it. There's very interesting ideas on both distortion and in room measurements.
I think a combination of many suggestions will work out extremely well.
As well as in post, as well as on the hardware side.
For example it still doesn't make any sense to use a omnidirectional microphone for measuring loudspeakers.
Just use a good cardioid microphone will already help significantly with reflections and other garbage.
Unfortunately those are a lot less common, and most aren't optimized for a linear frequency response, low (enough) THD and high dynamic range.
But yeah, it's kinda fascinating how sophisticated video call/stream apps are and how ancient old loudspeaker measurements software is.
That being said, in apps like Teams you only have to care about audibility, which is a lot easier than good total freq resp representation.
As well as in post, as well as on the hardware side.
For example it still doesn't make any sense to use a omnidirectional microphone for measuring loudspeakers.
Just use a good cardioid microphone will already help significantly with reflections and other garbage.
Unfortunately those are a lot less common, and most aren't optimized for a linear frequency response, low (enough) THD and high dynamic range.
But yeah, it's kinda fascinating how sophisticated video call/stream apps are and how ancient old loudspeaker measurements software is.
That being said, in apps like Teams you only have to care about audibility, which is a lot easier than good total freq resp representation.
My usage is a little different from what is normally meant by the term in mathematics.
What is done is to use that data in the impulse response that is without reflections and use that to extrapolate the impulse beyond the point of the first reflection by replacing the contaminated impulse response with the "linearly" extrapolated one.
There are many techniques for this and I did an AES preprint that reviewed them. The field is called Statistical Signal Processing. It uses linear regression to statistically analyze signals. You might want to look up the Prony Method or what are called ARMA (Auto-Regressive Moving Average) modeling. These techniques model the system as a polynomial where the coefficients are determined through linear regression. This subject is complex and highly mathematical so proceed with caution if your math skills are minimal.
Reflections in telephony are cancelled using ARMA models.
Hi @aslepekis
I'm currently working on a 3-way speaker project that requires a crossover from the bass drivers to the midrange driver around 250Hz. Unfortunately, the speaker is quite large and not easy to haul around inside/outside or even lift much to get longer gate windows.
The measurement files currently being utilized for crossover development were prepared using merged far-field (gates 4.0ms) + near-field (SPL-adjusted and Diffraction adjusted using VituixCAD Diffraction simulator). A crossover has been developed using those files, but I need to verify that it's working properly. I did take the speaker to a local community center, and got it off the ground to improve gate time from 4ms to 7ms (picture below) - but even 7ms isn't enough to verify performance in the 100-500hz range. The speakers are not easy to move - they weigh about 240lbs each and are approximately 64" tall. This is where I'm hoping the process being discussed in this thread could help.
I do not have a technical background (I'm a simple accountant - no crazy math or physics or accounting background) - but I'm hoping I can follow along.
I'm trying to confirm my understanding of the process outlined by you. So is it:
1) Take multiple measurements (minimum 8) over a distance of 6'-12' form DUT using ARTA or REW - single-channel mode
2) Export each impulse response as WAV
- Should I export in "32-bit Float" format?
- Should I be exporting "min phase version of IR"?
- Should I "normalize samples to peak value"?
4) Again using Audacity -- "Select" all tracks loaded, and then go to "Tracks" --> "Mix" --> "Mix and Render to New Track"
5) Select the newly mixed track and "File" -- > "Export Audio" -->Settings: Format: WAV, Export Range: Current Selection
6) In - REW ---> "File" --> "Import Impulse Response" - and open the file generated in step 5.
7) Gate the newly imported response to 30ms.
And, theoretically, what I should see after Step 7 would be a very close approximation to measurements obtained via-ground plane or NFS?
Is my understanding correct? Has this approach been superseded by your comments in #593 - which, (again, if I understood it correctly) -suggests taking measurements obtained in Step 1 above --> "Align SPL" --> " Vector Average"?
Im
