Equalizing loudspeakers based on anechoic measurements (community project)

[Maiky76]

@Maiky76

Would you care to generate filters for the KH80 and M105 based on your adjustment to the 8030C?

Hi,

Sure, my pleasure.
I wont publish curves, the scores and comments only in a spoiler to avoid bias as much as possible.
Also, I refrained from doing any manual tweaks, this is just the raw output of the optimizer.

format: Freq/Gain/Q

KH80
- no EQ score 6.15
- score with the following EQ:
EQ:
62.2, 1.33, 0.80,...
1992.0, 1.35, 1.62,...
5433.0, 0.77, 5.06,...
7830.0, 0.89, 5.78,...
9493.0, -0.55, 8.90,...
12990.0, -0.00, 7.30,...
23359.0, -2.70 1.37,..

M105
from the initial post, I think it is worth you time to close the loop on the EQ made by @amirm:
@amirm 4800.0, -2.00, 4.00,... score: 6.00
@amirm Optimized 4890.0, -2.20, 1.90,... score: 6.06
- score with the following EQ:

EQ:
300.5, -0.51, 3.05,...
299.0, 1.33, 0.94,...
1946.0, -0.94, 3.60,...
2917.0, 0.68, 2.45,...
5010.0, -1.63, 2.34,...
7991.0, 1.54, 1.13,...

The questions:
1. Is the EQ under test better or worst or indistinguishable from the unEQed speaker?
2. For the tested EQ which ones are better or worst or indistinguishable?

M

 

[Maiky76]

NBD_ON without EQ: 0.23
NBD_ON with EQ: 0.19

Interesting as well.
The NBD_ON from my EQ is 0.18, the NBD is heavily averaged so it is not very easy to hand tune the NBD.
It really is difficult to design actually, that's why I wrote the genetic algorithm based optimizer as opposed to an LMS (probably what REW is using) because I don't have a target to shoot for...

 

[TimVG]

M105
from the initial post, I think it is worth you time to close the loop on the EQ made by @amirm:
@amirm 4800.0, -2.00, 4.00,... score: 6.00
@amirm Optimized 4890.0, -2.20, 1.90,... score: 6.06
- score with the following EQ:

EQ:
300.5, -0.51, 3.05,...
299.0, 1.33, 0.94,...
1946.0, -0.94, 3.60,...
2917.0, 0.68, 2.45,...
5010.0, -1.63, 2.34,...
7991.0, 1.54, 1.13,...

The questions:
1. Is the EQ under test better or worst or indistinguishable from the unEQed speaker?
2. For the tested EQ which ones are better or worst or indistinguishable?

Testing the M105 first - sighted first (although I'm just uploading the EQ files without looking at the curves for now) -will do blind in the afternoon

As to answer the questions quickly

1. The single filters are ~somewhat~ noticeable but it depends on what is playing, and I may also be biased since it's a single filter and I know exactly what it's doing. Your filterset on the other hand is quite evident in an A/B switch.

2. The single filters may be better or indistinguishable (we'll know in the blind test) - I would say your generated set is better than the original M105. It seems at the same time politer and also clearer - I feel it sounds more natural. I'm comparing this filter to my own set at the moment (sighted still). Perhaps it were better if I were to simply use the uncorrected and compare blind to your set and my own.

Perhaps since we have another M105 owner @ElNino would like to participate in this test?

 

[flipflop]

The NBD_ON from my EQ is 0.18, the NBD is heavily averaged so it is not very easy to hand tune the NBD.

Last night I made a new EQ with a 0.8 dB slope and it also brought NBD_ON down to 0.18. The overall score remained at 6.9, though.

It really is difficult to design actually, that's why I wrote the genetic algorithm based optimizer as opposed to an LMS (probably what REW is using) because I don't have a target to shoot for...

I think your algorithm is very useful for determining the optimal slope. Could you please apply it to the newly measured GR Research X-LS and share the number it comes up with?

 

[Igor Kirkwood]

My english is poor . But I try send to ASR community 2 lasts curves of MMM mesurement by Jean-Luc Ohl of my Mastering Studio.
Level 100 dBC at the distance of listening 3,10 meters.
Notice that 4 subs are optimized for HC position with minimum L-R and better L+R

 

[TimVG]

In a blind test, I was able to pick out the factory M105 although it was more difficult than I'd imagined. Once again I had difficulty telling apart Maiky's filters (on the M105 this time) from my own, even though they were a little more apart this time, although not much (1dB).


Factory - my filters, then Maiky's

I feel as Qmuse noted, that we need a speaker with a directivity problem to further analyse the situation. (PIR vs LW/ON)

 

[TimVG]

@Maiky76 Do you need the Estimated In-Room .txt file for your calculation or will the ON, LW, ER and SP suffice? I have them for the F206 and C208 which are setup permanently and allow me to test for longer periods of time without cluttering up the house.

 

[Robbo99999]

In a blind test, I was able to pick out the factory M105 although it was more difficult than I'd imagined. Once again I had difficulty telling apart Maiky's filters (on the M105 this time) from my own, even though they were a little more apart this time, although not much (1dB).


Factory - my filters, then Maiky's

View attachment 75505

I feel as Qmuse noted, that we need a speaker with a directivity problem to further analyse the situation. (PIR vs LW/ON)

I don't think people would be able to tell between 2 different EQ curves if they're only 1dB apart, although actually it depends over how large an area of curve the 1dB difference applies to.....I remember when I was EQ'ing my speakers that a 1dB High Shelf from 1k (or was it 2k) was recognisable. So yes, I think it just depends on which part of the curves are affected by the 1dB difference and also over what sized area. It's interesting that the EQ curves being created here are so close together between the different people tweaking the EQ curves. I suppose we can only expect big/obvious differences if the EQ curves are more different from one another, but you can't make them intentionally bad of course (I mean why would you), so perhaps you guys are close to the zenith of what is achievable and are manipulating your EQ's within a really tiny window of overall effect.

 

[TimVG]

I don't think people would be able to tell between 2 different EQ curves if they're only 1dB apart, although actually it depends over how large an area of curve the 1dB difference applies to.....I remember when I was EQ'ing my speakers that a 1dB High Shelf from 1k (or was it 2k) was recognisable. So yes, I think it just depends on which part of the curves are affected by the 1dB difference and also over what sized area. It's interesting that the EQ curves being created here are so close together between the different people tweaking the EQ curves. I suppose we can only expect big/obvious differences if the EQ curves are more different from one another, but you can't make them intentionally bad of course (I mean why would you), so perhaps you guys are close to the zenith of what is achievable and are manipulating your EQ's within a really tiny window of overall effect.

What's interesting about Maiky's is he's using software to do to the EQ, while I'm 'eyeballing' it on mine. I feel we're definitely on the right track in any case.

 

[Robbo99999]

What's interesting about Maiky's is he's using software to do to the EQ, while I'm 'eyeballing' it on mine. I feel we're definitely on the right track in any case.

Well it's pretty cool that you both are very close together whilst using different processes to arrive at almost the same thing, albeit I suppose you're both taking the same variables into account. How much difference is there between your EQ that you've managed to do from the knowledge of the spinorama vs the best that you were able to achieve before spinorama knowledge? To me that's an interesting point too, because without spinorama you don't have all the info and it's harder to EQ the speaker rather than the room so to speak. I mean that's why I'm looking forward the JBL 308p review that Amir is gonna be putting out at some point, because I'm hoping it will allow me to apply more relevant EQ than UMIK in room measurements alone. (I think I've inferred what the spinorama is gonna look like for on-axis from my many UMIK measurements, in terms of major characteristics/flaws, but will still be good to get confirmation on that as well as some additional detail.)

 

[TimVG]

How much difference is there between your EQ that you've managed to do from the knowledge of the spinorama vs the best that you were able to achieve before spinorama knowledge?

It's just so much easier because you have a much more complete and accurate picture of what is going on. With Amir's data set especially anything can be extracted (pretty much) in terms of amplitude response.

Thanks to a tip from aarons915 to use Vituixcad it's even possible to plot spinorama's from image files and import them into REW which I've been using for my F206 and C208 - it works very well.

 

[Robbo99999]

It's just so much easier because you have a much more complete and accurate picture of what is going on. With Amir's data set especially anything can be extracted (pretty much) in terms of amplitude response.

Thanks to a tip from aarons915 to use Vituixcad it's even possible to plot spinorama's from image files and import them into REW which I've been using for my F206 and C208 - it works very well.

A quick question on importing spinorama data. I had a look at the data file that Amir attaches to his reviews. I managed to extract them to text files, and it looks like I could just import some of them directly into REW for manipulation although I've not needed to try that yet. I figured I might need to cut & paste the relevant columns and put them into a new text file if worst comes to the worst. Does that sound sensible, and is there any corrections that need to be done to the actual data included in those files or can we use the actual data as is....I remember some folks talking about 'corrections' but I'm probably getting my wires crossed?

 

[bobbooo]

Allow me to present a different perspective: r² is defined as 1-(SSres/SStot). If you take a given PIR curve, and you increase its slope (but keep local deviations the same), the SSres of the (newly recomputed) regression will stay the same, but SStot will increase (because PIR points are moving farther away from the mean). The result: r², i.e. SM, increases. Conclusion: local deviations being kept the same, if the overall tilt of the PIR curve increases, SM_PIR will increase.

Do you mean if the overall magnitude of the PIR slope increases? If so, this means the formula (incorrectly) rewards extreme PIR slopes, both bright and dark (positive and negative). In practice there will likely be very few speakers with a PIR slope above 0, so the excessively bright scenario is likely less of a problem, but the overly dark scenario might be more common. I think this all means it's vital to do a 'sanity check' on each speaker's computed score by cross-referencing it with its PIR slope value, which should be close to the 'ideal' -1 value. Alternatively, as there is a slight discrepancy between the ideal PIR slope of the bookshelf (-1.2) and the 'all-speaker' (-1) tests, which Olive suggests could be down to the latter on average having wider dispersion, maybe the best slope value to look at would simply be the on-axis, which for both tests have an ideal value of exactly 0. (A further cross-check could even be done with the listening window slope, which again agree on an ideal value, -0.2, between the two tests.)

For EQing, I think the best approach would then be to optimize for the highest Olive score, but only if the slope of the (in order of importance) on-axis / listening window / PIR move further toward their 'ideal' values, or stay the same. Otherwise the above-mentioned shortcoming of the Olive formula might be 'gamed' for a higher score at the expense of worse overall tonality. @Maiky76 is this something that could be implemented in your ideal EQ computation?

 

[Robbo99999]

Do you mean if the overall magnitude of the PIR slope increases? If so, this means the formula (incorrectly) rewards extreme PIR slopes, both bright and dark (positive and negative). In practice there will likely be very few speakers with a PIR slope above 0, so the excessively bright scenario is likely less of a problem, but the overly dark scenario might be more common. I think this all means it's vital to do a 'sanity check' on each speaker's computed score by cross-referencing it with its PIR slope value, which should be close to the 'ideal' -1 value. Alternatively, as there is a slight discrepancy between the ideal PIR slope of the bookshelf (-1.2) and the 'all-speaker' (-1) tests, which Olive suggests could be down to the latter on average having wider dispersion, maybe the best slope value to look at would simply be the on-axis, which for both tests have an ideal value of exactly 0. (A further cross-check could even be done with the listening window slope, which again agree on an ideal value, -0.2, between the two tests.)

For EQing, I think the best approach would then be to optimize for the highest Olive score, but only if the slope of the (in order of importance) on-axis / listening window / PIR move further toward their 'ideal' values, or stay the same. Otherwise the above-mentioned shortcoming of the Olive formula might be 'gamed' for a higher score at the expense of worse overall tonality. @Maiky76 is this something that could be implemented in your ideal EQ computation?

Yes, my intuition is telling me to EQ on-axis, especially if directivity is good, and if you toe in speakers pointing right at your listening position. That's gonna be my first approach once Amir reviews the JBL 308p Mkii.

 

[TimVG]

Yes, my intuition is telling me to EQ on-axis, especially if directivity is good, and if you toe in speakers pointing right at your listening position. That's gonna be my first approach once Amir reviews the JBL 308p Mkii.

What seems to work well is EQ'ing for a flat(tish) listening window, subsequently applying the filters to the on-axis response, and using the appropriate shelf to make sure there is no excess energy in the on-axis curve. This way the narrow band EQ filters are only applied on the listening window average and eliminates the chance that you are EQ'ing on-axis interference effects which are not meant to be EQ'd.

When this is done the filters can be applied to the early reflections (or in-room estimate) curve and see if nothing too weird is going there. Faults of omission are generally considered more benign than faults of excess.

Here's my F206s

 

[TimVG]

A quick question on importing spinorama data. I had a look at the data file that Amir attaches to his reviews. I managed to extract them to text files, and it looks like I could just import some of them directly into REW for manipulation although I've not needed to try that yet. I figured I might need to cut & paste the relevant columns and put them into a new text file if worst comes to the worst. Does that sound sensible, and is there any corrections that need to be done to the actual data included in those files or can we use the actual data as is....I remember some folks talking about 'corrections' but I'm probably getting my wires crossed?

Yes, I also cut the relevant collumns and make a new txt file. I'm not sure if any further corrections need to be performed, perhaps the microphone cal file needs to be embedded, which I hadn't considered up until now - perhaps someone can provide the answer.

 

[edechamps]

Do you mean if the overall magnitude of the PIR slope increases?

Yep, that's what I meant, strictly speaking. Thank you for clarifying.

If so, this means the formula (incorrectly) rewards extreme PIR slopes, both bright and dark (positive and negative).

Yep, exactly. It's part of the fun of Olive misusing r² for his SM metric. It's the gift that keeps on giving.

maybe the best slope value to look at would simply be the on-axis, which for both tests have an ideal value of exactly 0. (A further cross-check could even be done with the listening window slope, which again agree on an ideal value, -0.2, between the two tests.)

I agree that targetting an ON (or LW) slope of 0 makes more sense than targetting an "ideal PIR slope", which is a concept I always found dubious. Some people might prefer a different slope, but that's fine - they can always apply their own preferences on top of a "flat" EQ.

the above-mentioned shortcoming of the Olive formula might be 'gamed' for a higher score at the expense of worse overall tonality.

This is my main concern about this whole "EQ based on score" project. The Olive model was designed to correlate with preference on a typical sample of loudspeakers. Just because something correlates with good sound in a typical speaker doesn't mean that it causes good sound, and for that reason, using the model as an optimization target might be missing the point entirely. (Though I will gladly concede that it's likely better than doing nothing.)

If I wanted to EQ my speaker based on anechoic measurements, I think I would simply EQ the LW to flat and call it a day. Any remaining problems would be caused by uncontrolled directivity, which can't be fixed with EQ anyway.

 

[Robbo99999]

What seems to work well is EQ'ing for a flat(tish) listening window, subsequently applying the filters to the on-axis response, and using the appropriate shelf to make sure there is no excess energy in the on-axis curve. This way the narrow band EQ filters are only applied on the listening window average and eliminates the chance that you are EQ'ing on-axis interference effects which are not meant to be EQ'd.

When this is done the filters can be applied to the early reflections (or in-room estimate) curve and see if nothing too weird is going there. Faults of omission are generally considered more benign than faults of excess.

Here's my F206s

View attachment 75583
View attachment 75584

Ah, ok, so to rephrase what you're saying to make sure I understand properly.

1. Use narrow band filters to EQ the listening window to an un-rippled line (so it's smoother).
2. Use broad tone control EQ on top of that to EQ the on-axis to level (but not flat, ie not un-rippled, it can remain a little jagged).
3. Add all these filter to the In Room Estimate curve to check it doesn't do anything too crazy...because I guess that would be the equivalent of using UMIK to actually measure the EQ results in our own rooms, minus room modes of course.
4. If it passes those 3 prior tests/steps, then apply EQ to speakers & measure using UMIK in our own rooms, apply room EQ to sub 500Hz.

How's that, is that correct?


EDIT: I just thought of a possible Step #5: Consolodate all those layers of filters into simpler filters by charting the final EQ curve as a target and then using REW to EQ a flat 0dB line to the EQ Curve target using the smallest number of filters possible. This might get rid of conflicting filters. And potentially it can sound better with fewer filters right, even if it ends up with the same final EQ curve??

 

[flipflop]

@edechamps
The problem with basing the EQ on the on-axis or listening window is that certain speakers with recessed treble and strong early reflections will have their SM_PIR degraded to the point where it drags down the entire aggregated score.
Equalizing to the PIR is the only universal approach that works.

 

[TimVG]

Ah, ok, so to rephrase what you're saying to make sure I understand properly.

1. Use narrow band filters to EQ the listening window to an un-rippled line (so it's smoother).
2. Use broad tone control EQ on top of that to EQ the on-axis to level (but not flat, ie not un-rippled, it can remain a little jagged).
3. Add all these filter to the In Room Estimate curve to check it doesn't do anything too crazy...because I guess that would be the equivalent of using UMIK to actually measure the EQ results in our own rooms, minus room modes of course.
4. If it passes those 3 prior tests/steps, then apply EQ to speakers & measure using UMIK in our own rooms, apply room EQ to sub 500Hz.

How's that, is that correct?

That's spot on and what's giving me the best listening experience so far. I'd like to add that depending on how you're measuring and what the distance and room qualities are like, the predicted slope may not match the measured slope, especially in the treble. Going by the estimate does give us an 'untainted' picture that doesn't include potential measurement artifacts (let's say a reflection from the back of the couch).