Equalizing loudspeakers based on anechoic measurements (community project)

[TimVG]

Hi Guys (& gals?)

Since there's been some discussion about how the different anechoic curves affect listener preference I'm going to attempt to gather some people here who own loudspeakers that have been tested here on ASR. The goal of the test is simple. We're going to experiment with different filter settings based on Amir's anechoic measurements and see what we come up with.

Some context: There have been many speculations about what is most important to us as listeners. Some claim the listening window needs to be as flat as possible, others want the estimated in-room response to be as smooth as possible, etc.
I'm in luck as to owning three (well designed) models that have also been tested by Amir, and have been experimenting quite a bit over the last few weeks if there is indeed something that can be universally applied to improve most if not all loudspeakers **please note I'm fully aware there are many things EQ can not fix**

I'm now looking for volunteers who also own speakers tested by Amir, to casually try some filter settings. If you can do it blind, all the better, but any input is appreciated.

Prerequisites:
A good start would be that your system is already calibrated in the low range. Please note that any filter settings based on anechoic measurements that are in your own standard 'EQ range' will be superimposed on said filters and will likely further affect perception. They should be turned off or adjusted so that the low range remains constant. If you can share some details about the position of your speakers relative to the walls this would also be appreciated.

TBC

 

[TimVG]

Targets:

Below we have a comparison of 4 different slopes of the Revel M105

The above set being the on-axis, listening window and estimated in-room response as-is from the factory.

The second set has filters optimised for a flat listening window. As we can see this has a consequences for the on-axis sound which now has some peaking introduced as well as some excess energy in the in-room estimate.

The third set is optmised on the predicted in room response following a -1dB slope per octave from 20Hz to 20kHz. The on-axis peaking is now eliminated and the and the listening window remains relatively smooth.

The fourth set, and my favourite, is quite close to the third set in terms of result,but introduces back some of the on-axis energy lost when concentrating solely on the estimated in-room curve (which as we will see in different loudspeakers is quite dependent on overall directivity. Interestingly the estimated in-room response is a happy accident as I did not focus on it when implementing the filters. The idea was to keep the on-axis flat in terms of general trend, and balance the listening window and sidewall reflection (not shown) with this in mind.

comparison of on-axis / listening window of filterset 3 and 4

 

[TimVG]

The same data for the Neumann KH80

In a more detailed view: when optimising for the listening window we can see peaking developing in the on-axis sound

And when optimising for the in-room estimate we can see clear difference in the direct sound where normally we normally have a flat response. By trying to force the in-room estimate to a set curve this causes issues: an excess of energy from 1kHz onward, followed by too little after 8kHz. This is because we are trying to force a downward sloping curve onto an otherwise constant directivity design. It doesn't work very well.

 

[TimVG]

Same data for the Genelec 8030C

Optimising for the listening window only again introduces an excess amount of energy on-axis

If we target a 'perfect' in-room response:

Unfortunately this compromises the direct sound to quite an audible extent.

 

[TimVG]

My own conclusions so far:

On axis response needs to be free of excess amounts of energy and the general slope needs to be flat. There cannot be an up- or downward trend. Some discrepancies however appear to be audibly inoffensive (Revel performa3 seems to come to mind) The listening window must be relatively smooth with a neutral to slightly downward sloping trend - if correcting it means implementing a trend in the on-axis curve, don't.
Predicted in-room response will be directivity dependent and a fixed target is wrong.

More work needs to be performed, working with loudspeakers that have directivity that is perhaps less good.

 

[TimVG]

One so far unmentioned aspect of subjective differences is the early sidewall reflection.
To quote Revel's Kevin Voecks

"Harman research has discovered that the first reflection from side walls, both from the wall adjacent as well as the opposite side wall are critically important. The acoustic output of a loudspeaker far off-axis horizontally is very significant, and should match the response of the Listening Window as much as possible. This goal is technically challenging, but is essential for optimum timbre, as well as to provide a sense of seamless coherency."

Let's take a look shall we.

Here are the sidewall reflection curves for the Genelec 8030C, Neumann KH80 and the Revel M105 (after correction)

Clearly there are some differences here. To see which model matches the response of the listening window as much as possible, we'll calculate a directivity index based on said listening window.

8030C

KH80

Revel M105

Sidewall directivity overlay

These measurements could help explain some of the subjective differences between loudspeakers that otherwise offer good measurements before traversing into different areas such as distortion.

 

[QMuse]

Loudspeaker EQ strategy should be derived from the target you want to achieve and that target is based on listener preference scoring system. If your target is to EQ a speaker to achieve max Olive's score than you can derive the EQ strategy from the components that make the score and their weighting in the Olive's score.

If you want to EQ to different target by, for example taking into account horizontal directivity and giving some weight to it, you should first be sure that such modified scoring system better matches listener preference than Olive's model and then you can derive EQ target from that new scoring system.

In any case you don't start from the EQ strategy but from the preference scoring system as EQ strategy is derived from it.

 

[TimVG]

I don't consider the preference scoring system to be absolute. While the science behind it is correct, the test(s) involved too many bad and mediocre loudspeakers. In testing 3 loudspeakers that would end up with high score and close to a statistical tie, there are unmistable audible differences. Please note I'm not talking about preference now, merely differences. Also the initial goal was met, which was to improve each model based on anechoic measurements.

If you'd like, you could generate me some EQ files for each of the models optmised according to the preference score? I'm planning on testing the EQ files blind soon anyhow.

 

[QMuse]

I don't consider the preference scoring system to be absolute. While the science is correct, the test(s) involved too many bad and mediocre loudspeakers. In testing 3 loudspeakers that would end up with high score and close to a statistical tie, there are unmistable audible differences. Please note I'm not talking about preference now, merely differences. Also the initial goal was met, which was to improve each model based on anechoic measurements.

If you'd like, you could generate me some EQ files for each of the models optmised according to the preference score? I'm planning on testing the EQ files blind soon anyhow.

My point was that optimal EQ strategy is directly derived from the preference score and Olive's preference score is, at least to my knowledge, the best one that exists. In order to have better EQ strategy you need better preference scoring system. So, you don't start from changing EQ strategy, you start from changing preference scoring system as optimal EQ strategy is based on it. But changing preference scoring system is not a trivial thing to do as for any change, like for example putting some weight on the horizontal directivity, you would need to check if it really better matches listener preference, which is complicated and expensive.

As long as Olive's preference scoring is the best one we have it is wise to EQ the speaker to achieve highest Olive's score as any other EQ strategy is just a shot into the dark.

 

[TimVG]

My point was that optimal EQ strategy is directly derived from the preference score and Olive's preference score is, at least to my knowledge, the best one that exists. In order to have better EQ strategy you need better preference scoring system. So, you don't start from changing EQ strategy, you start from changing preference scoring system as optimal EQ strategy is based on it. But changing preference scoring system is not a trivial thing to do as for any change, like for example putting some weight on the horizontal directivity, you would need to check if it really better matches listener preference, which is complicated and expensive.

As long as Olive's preference scoring is the best one we have it is wise to EQ the speaker to achieve highest Olive's score as any other EQ strategy is just a shot into the dark.

I understand your point, doing nothing however won't help anything either, so unless you're willing to provide me with some EQ files optimised according to the preference rating system for me to compare blind, or volunteer to test some files yourself - there isn't much more to add.

 

[QMuse]

I understand your point, doing nothing however won't help anything either, so unless you're willing to provide me with some EQ files optimised according to the preference rating system for me to compare blind, or volunteer to test some files yourself - there isn't much more to add.

To EQ a speaker based on it's spinorama measurement to achieve highest Olive's preference score you would need a software designed to do just that, and no such software exists to my knowledge. What I'm aming for when doing speaker EQ is aiming equally for flat LW and smooth PIR as I think they are equally important. I'm not really claiming that is the best EQ strategy but it will get you close to the highest possible score and will come close to producing the best tonal balance particular speaker can theoretically achieve.

 

[TimVG]

To EQ a speaker based on it's spinorama measurement to achieve highest Olive's preference score you would need a software designed to do just that, and no such software exists to my knowledge. What I'm aming for when doing speaker EQ is aiming equally for flat LW and smooth PIR as I think they are equally important. I'm not really claiming that is the best EQ strategy but it will get you close to the highest possible score and will come close to producing the best tonal balance particular speaker can theoretically achieve.

The PIR is dictated by directivity vs lineair behaviour in the direct sound - for certain designs it's simply impossible to have both (constant directivity systems come to mind) - would that mean it would be less preferred assuming both have an equally good on-axis/LW? Maybe.. the Salon2 did beat out the M2 that one time. It are all interesting questions yet to be answered. My latest EQ scheme is not much different from yours with the main difference that I don't allow for on-axis peaking as shown above.

 

[QMuse]

The PIR is dictated by directivity vs lineair behaviour in the direct sound - for certain designs it's simply impossible to have both (constant directivity systems come to mind) - would that mean it would be less preferred assuming both have an equally good on-axis/LW? Maybe..

Given equal on-axis/LW a speaker with smoother PIR would be preferred, at least acoording to Olive's score.

When you can't have both, LW and PIR, smooth because of directivity errors, I would try to make LW smooth with short listening distances and PIR smoth with long listening distances. With mid-range listening (say between 1.5m and 2.5m) I would divide the smoothness errors equally between them.

the Salon2 did beat out the M2 that one time.

Hard to say why, but if I would have to bet I would put my money that Salon2 won because of wider horizontal directivity.

My latest EQ scheme is not much different from yours with the main difference that I don't allow for on-axis peaking as shown above.

Yes, my logic is basically very similar to yours. I put more value in LW than on on-axis because on-axis is a single measurement while LW is a spatial average. What I would like to see is LW measured in a narrower range than what it is now, say average of 0, 5, 10 and 15 deg. IMO that would better serve as direct sound representative for EQ than current LW and/or on-axis.

 

[TimVG]

Yes, my logic is basically very similar to yours. I put more value on LW than on on-axis as on-axis is a single measurement. What I would like to see is LW measured in a narrower range than what it is now, say average of 0, 5, 10 and 15 deg. IMO that would better serve as direct sound representative for EQ than current LW and/or on-axis.

I too put more value on the LW in terms of smoothness and narrow band deviations, but in terms of relative trend I prefer to keep the on-axis flat as shown in the examples above.

When you can't have both, LW and PIR, smooth because of directivity errors, I would try to make LW smooth with short listening distances and PIR smoth with long listening distances. With mid-range listening (say between 1.5m and 2.5m) I would divide the smoothness errors equally between them.

So far with Genelecs at least, my EQ preference doesn't seem to change with distance (in my treated room at 2m or untreated living room at 3,5m). I plan to repeat this test blind (someone else will upload the filtersets).

 

[QMuse]

I too put more value on the LW in terms of smoothness and narrow band deviations, but in terms of relative trend I prefer to keep the on-axis flat as shown in the examples above.

To be honest I would never buy a speaker which LW and on-axis differ significantly.

So far with Genelecs at least, my EQ preference doesn't seem to change with distance (in my treated room at 2m or untreated living room at 3,5m). I plan to repeat this test blind (someone else will upload the filtersets).

Genelecs are good wherever you put them. When testing such things IMO you would better do with a speaker that has directivity error(s), such as that Salk that was recently tested. With it you would have to choose either to have flat LW or smooth PIR because of that dip with DIs at XO region.

 

[TimVG]

To be honest I would never buy a speaker which LW and on-axis differ significantly.

Sure, however if you look at the Genelec and Neumann examples above you'll see the on-axis and LW are actually very close to eachother - if you however EQ for a completely flat LW there is a small rising response issue in the on-axis - this to me is audible, but we'll see how I end up in the blind test.

Genelecs are good wherever you put them. When testing such things IMO you would better do with a speaker that has directivity error(s), such as that Salk that was recently tested. With it you would have to choose either to have flat LW or smooth PIR because of that dip with DIs at XO region.

Yes, that's why I'm looking for volunteers.

 

[aarons915]

My own conclusions so far:

On axis response needs to be free of excess amounts of energy and the general slope needs to be flat. There cannot be an up- or downward trend. Some discrepancies however appear to be audibly inoffensive (Revel performa3 seems to come to mind) The listening window must be relatively smooth with a neutral to slightly downward sloping trend - if correcting it means implementing a trend in the on-axis curve, don't.
Predicted in-room response will be directivity dependent and a fixed target is wrong.

More work needs to be performed, working with loudspeakers that have directivity that is perhaps less good.

You know I've been interested in this topic for a long time as well and the ideal off-axis response. One thing I think is important is when you're this into EQ you can almost make your own personal curves, for example I don't personally pay attention to the direct on-axis curve because I don't toe my speakers at all so my direct sound is around 20 degrees. Also I personally use the ER curve but I see you like to separate the sidewall reflections, since you place more importance on it I would make sure it matches your actual setup, I know in my room with my speakers having no toe, my sidewall reflection is right around 45 degrees so I actually use the average of 40 and 50 deg when looking at that. It's a lot of work and I'm not that obsessed (yet) but you could really create your own personal CTA-2034 graph based on your own room and setup if you really wanted to, remember the spec is based on averages that may or may not be close to your situation.

I think you're on to something about the PIR slope having to do with the speaker directivity and also the size of the room, I don't worry about the actual in room response even though this does make it better as well. My preferred method is to basically only look at the LW and ER curves and make a compromise between the 2 but based on about a month of testing the R3 I've decided on a filter set where the ER curve is a gently falling slope shown here:

I'd be really curious on your impressions on the Neumann and Genelec, they are both very neutral but different ER curves, what kind of differences do you notice comparing the 2?

 

[TimVG]

You know I've been interested in this topic for a long time as well and the ideal off-axis response. One thing I think is important is when you're this into EQ you can almost make your own personal curves, for example I don't personally pay attention to the direct on-axis curve because I don't toe my speakers at all so my direct sound is around 20 degrees. Also I personally use the ER curve but I see you like to separate the sidewall reflections, since you place more importance on it I would make sure it matches your actual setup, I know in my room with my speakers having no toe, my sidewall reflection is right around 45 degrees so I actually use the average of 40 and 50 deg when looking at that. It's a lot of work and I'm not that obsessed (yet) but you could really create your own personal CTA-2034 graph based on your own room and setup if you really wanted to, remember the spec is based on averages that may or may not be close to your situation.

I think you're on to something about the PIR slope having to do with the speaker directivity and also the size of the room, I don't worry about the actual in room response even though this does make it better as well. My preferred method is to basically only look at the LW and ER curves and make a compromise between the 2 but based on about a month of testing the R3 I've decided on a filter set where the ER curve is a gently falling slope shown

That also potentially explains why you thought my R3 filter sounded just a tad more recessed (from our AVS conversation) than your own - Don't worry I'm not going to ask you to toe them in.
I do like to keep an eye at the side wall reflection, but it's not that I put more emphasis on it (anymore). Usually when correcting the other curves, the sidewall curve follows, and vice versa. I tend to end up with a balance as you all seem to. I also don't mind the measured in-room response, but the anechoic prediction does seem to have its place.


I'm going to crossquote you from AVS here for a second

So I plugged your EQ into a comparison file I use to see the differences and I don't know if this was intentional but your EQ'd listening window and ER curves match the Revel Be line almost exactly, most notably the F226be. I've looked at many of that lines curves myself because to me it should show the best sound possible at the moment from their latest research.

I had not noticed that. I just uploaded the F328be curves - this is almost exactly what I'm aiming for. No peaking in the on-axis, relatively smooth/neutral LW and a relatively clean PIR. There's actually very little to EQ here, at least in view of what I'm aiming for, half a dB here and there - that's it! Remarkable for a passive loudspeaker.

I'd be really curious on your impressions on the Neumann and Genelec, they are both very neutral but different ER curves, what kind of differences do you notice comparing the 2?

Listening in mono doesn't flatter the Neumann in a direct comparison, and I mean this is a benign way - if you check out the sidewall curve (this is an example of why it's useful) you can see although well behaved it features a very downward aimed slope. It's less 'enveloping' than the Genelec or the Revel. In stereo however this comes across as more focussed (? if that's the word). Timbrally speaking they are quite close - as the on-axis/LW curves indicate they do.

 

[aarons915]

That also potentially explains why you thought my R3 filter sounded just a tad more recessed (from our AVS conversation) than your own - Don't worry I'm not going to ask you to toe them in.
I do like to keep an eye at the side wall reflection, but it's not that I put more emphasis on it (anymore). Usually when correcting the other curves, the sidewall curve follows, and vice versa. I tend to end up with a balance as you all seem to. I also don't mind the measured in-room response, but the anechoic prediction does seem to have its place.

It definitely contributes but I think generally not toeing your speakers in is going to make the direct sound and sidewall reflection more similar just due to the fact that the angles are now much more similar, roughly 20 degrees to 45 instead of 0 and closer to 75 degrees, so I prefer to just start with that in mind. Also regarding the F328be, it does measure great for a passive like most Revel but I have to think based on my listening to that similar EQ, most would find them just a bit recessed in that range, I think a constant directivity design wants to have an ER curve right about 3 db below the listening window based on the comparisons I've done. Not too surprisingly, this is basically how the Salon 2 measure and most think they are about perfect. One positive about that slight recessed sound is that you can really crank them and the highs never bother me, while my stronger but declining ER curve are a bit energetic sounding at loud volumes.

Listening in mono doesn't flatter the Neumann in a direct comparison, and I mean this is a benign way - if you check out the sidewall curve (this is an example of why it's useful) you can see although well behaved it features a very downward aimed slope. It's less 'enveloping' than the Genelec or the Revel. In stereo however this comes across as more focussed (? if that's the word). Timbrally speaking they are quite close - as the on-axis/LW curves indicate they do.

Have you done a carefully controlled and level-matched A/B between the 2 where you could instantly switch between the 2? I found comparing our EQ that sometimes it was hard to tell a difference, it would be impossible if I couldn't instantly switch back and forth. I also found the most notable differences to be female vocals, violin and the banjo. I know there doesn't seem to be a consensus on what the ideal off-axis response is but since the ER curve so closely resembles the PIR and apparently the ideal room curve is smoothly declining, it only makes sense that the same has to be true for the ER curve.

 

[Maiky76]

To EQ a speaker based on it's spinorama measurement to achieve highest Olive's preference score you would need a software designed to do just that, and no such software exists to my knowledge. What I'm aming for when doing speaker EQ is aiming equally for flat LW and smooth PIR as I think they are equally important. I'm not really claiming that is the best EQ strategy but it will get you close to the highest possible score and will come close to producing the best tonal balance particular speaker can theoretically achieve.

Hi,

I have just such a SW running in Matlab.
I have been publishing EQ suggestions based on Score-optimized EQ:

https://www.audiosciencereview.com/...ctive-speaker-review.13436/page-3#post-432075
https://www.audiosciencereview.com/...tudio-monitor-review.14795/page-5#post-461044
https://www.audiosciencereview.com/...-m16-speaker-review.11884/page-23#post-458542
https://www.audiosciencereview.com/...shelf-speaker-review.14745/page-2#post-458448


I agree with you that the only known/documented model is the Olive score and therefore we should stick with it.
Any deviation from it would need redoing the whole research for validation.

On the other hand my observations are very similar to TimVG's:
- with a speaker with partially constant directivity (i.e flat SPDI at HF, with a waveguide on the HF unit for example, non-coaxial ) the score improves by tilting down the whole response otherwise the PIR HF exhibits too much energy. This is a very important observation I believe.
- the midrange gets boosted to compensate (fill the lack of energy) for the directivity error around the crossover.

The following is based on idealized models so please no nitpicking I am just trying to extract trends…

The preference score Equation is:
PPR_ON = 12.69 - 2.49*NBD_ON - 2.99*NBD_PIR - 4.31*LFX + 2.32*SM_PIR;

If one ignores the LFX to avoid over-stretching the speaker, or latter on add multiple LF sources, then the score driving factor is the PIR.
One should therefore design the speaker with the PIR as priority.

This is linked with the research based on the in-room EQ AND the headphone target curve research both conducted by Harman.

“My” Interpretation:
- 2.49*NBD_ON (100Hz-12000Hz): the score benefits from a flat response with as little deviations from the average value as possible; we want NBD_ON=0.
Even adding a gentle slope would not result asymptotically in NBD_ON=0.
However, having the curve that looks like some stairs on each 1/2 octave band could yield 0 as well; not that I advocate this kind of design it is just a remark…

- 2.99*NBD_PIR (100Hz-12000Hz) same as NBD_ON, we want NBD_PIR=0
- 4.31*LFX: we want 4.31*LFX = 0 which means a 6dB cut-off of the SP to be 1Hz relative to the average of the ON response in the 300-10000Hz range.
If one wanted max Score = 10 (not sure why) then the LFX could be f6dB = 14.5Hz.
+ 2.32*SM_PIR (100-16000Hz) is comprised between 0 and 1, we want SM_PIR = 1.
To calculate this property of the PIR, first, one needs to make a LINEAR regression of the PIR, and then, the SM expresses the validity of this regression.
The more the PIR resembles the LINEAR regression the closer to 1 the SM_PIR will be. The PIR should therefore be a LINE with a slope, ANY slope.
This eliminates the stairs “case” from the NBD, as, in all likelihood, the derived PIR would not resemble a line.



This leads to design considerations on the target one should follow to achieve the best score possible.

• NBD_ON = 0 translates into Flat for ON, OK nothing new there.
• LFX: 14.5Hz, more reasonable that 1Hz… Doable with dedicated SWs stand alone or not and EQ
• SM_PIR = 1 means PIR is a line with with a slope, ANY slope. That is the crux of the matter I believe.

Now, knowing that we also want NBD_PIR = 0 it means that the PIR should therefore be a FLAT line i.e. slope = 0.

• NBD_PIR = 0 translates into Flat for PIR,

Remember PIR = 0.12*LW + 0.44*ER + 0.44*SP

Then a flat PIR also “probably” means

• Flat LW
• Flat ER
• Flat SP

So now we have the “idealized” targets:

• Flat ON: not a surprise
• Flat LW: no variation on the tonal balance across the LW, not a surprise
• Flat ER: no variation on the tonal balance with the ER, not a surprise
• Flat SP: consequence of the rest of the targets, maybe not realistic, the SP contribution is much lower than the ER and LW in the PIR calculation so less critical

Now how does one make such a speaker, at least on the horizontal plan?

• Controlled and constant directivity down to 100Hz via large drivers, waveguides and/or beam forming

https://www.stereophile.com/content/bang-olufsen-beolab-90-loudspeaker-measurements
https://www.stereophile.com/content/dutch-dutch-8c-active-loudspeaker-system-measurements
https://www.stereophile.com/content/kii-audio-three-loudspeaker-measurements
https://www.audiosciencereview.com/forum/index.php?threads/apple-homepod-measurement.8425/

Another good approximation:
https://3.bp.blogspot.com/-OB4hm25dXms/XJVUs8TznTI/AAAAAAAAAEA/r6riUCqhZDgJO61yR8uKzWXbHDYsl_CJgCLcBGAs/s1600/Spin+-+Revel+Performa3Be+F228Be.png


That is also the target for Earl Geddes if I am not mistaken the difference being the DI value more of less high i.e. refection contribution.

• Onmi-directional is a special case of this that might requires multiple HF sources to extend the omni character up to HF

https://www.audiosciencereview.com/...directivity-speaker-review.13982/#post-426504

This is a very old debate that I do not pretend to solve here but that seems to tilt the balance towards constant directivity (Flat SPDI) vs monotonic increasing directivity i.e. a SPDI that increases constantly with frequency.
The latter providing:

• Flat ON, OK for NBD_ON = 0
• Tilted LW with built in issue with NBD_PIR = 0
• Tilted ER with built in issue with NBD_PIR = 0
• Tilted SP that might not be an issue with reasonably directive speakers (high DI)
• Tilted PIR with built in issue with NBD_PIR = 0 but with no reason NOT to achieve SM_PIR = 1 target

Example:
https://speakerdata2034.blogspot.com/2019/03/spinorama-data-kef.html

With all that been said, the score relies on anechoic data that does not include the room influence per definition.
So I guess one way to see things would be to stick to the these anechoic data within the limits of room/speaker dominated domains
Therefore we could restrict the target to the Schroder frequency (which one?) 500Hz as most large speaker Summa, JBL M2 or ?
to get a reasonable approximation of the design targets.


Sorry for the long post but I thought it was the right time to expand on my thoughts.

Cheers
M