Equalizing loudspeakers based on anechoic measurements (community project)

[Chromatischism]

As I mentioned before, hasn't it been proved that our brain/ears cannot distinguish between direct sound and early reflections arriving in first X ms?

If that is so than information contained in ER and SP are added to direct sound and perceived joined. As PIR mostly consists of ER and SP (and partly of direct sound) there lies the basis for using PIR as an efficient EQ strategy.

The problem is that PIR penalizes very wide dispersion speakers, and the subjective impressions diverge from what the model suggests. I'm sure you've seen the Philharmonic BMR reviews.

 

[TimVG]

The problem is that PIR penalizes very wide dispersion speakers, and the subjective impressions diverge from what the model suggests. I'm sure you've seen the Philharmonic BMR reviews.

Actually if we EQ that loudspeaker to feature a reasonably flat direct sound, that PIR is looking pretty 'textbook'

Maybe the limited vertical dispersion balances it out by chance. Some designs are definitely penalized in terms of score though - with or without merit.. that's a big debate.

 

[Chromatischism]

Actually if we EQ that loudspeaker to feature a reasonably flat direct sound, that PIR is looking pretty 'textbook'

View attachment 75988

What I'm referring to is that the PIR suggested that it would be a bright sounding speaker. It's revealing, but not really bright, according to every review I've read.

That looks like a good result.

 

[TimVG]

What I'm referring to is that the PIR suggested that it would be a bright sounding speaker. It's revealing, but not really bright, according to every review I've read.

That looks like a good result.

Could also be attributed to it not as neutral as possible (although it's nothing problematic), The wide horizontal dispersion which doesn't really show up in the ER/PIR. If you isolate the sidewall reflection however it clearly shows. I'd say certain constant directivity designs are penalized more so than this loudspeaker. I'm willing to bet the GGNTKT M1 wouldn't score as high as some others, yet a constant directivity down to the room transition would seem in fact desirable as the horizontal reflected sound would match the direct sound extremely well in just about all of the frequency range where it actually matters.

 

[JIM_82]

I think maybe there's some confusion here.

There are two ways to interpret PIR: what it's intended to model, i.e. an estimate of how a typical in-room curve might look like; and what it's defined as (i.e. how it's computed), which is just a weighted average of various curves: 12% Listening Window, 44% Early Reflections, 44% Sound Power. In a way, you could see it as a variation on Sound Power where some of the curves have more weight than others.

But in both cases, PIR is not representative of each particular case.
See the graph above on the direct and diffuse distribution, it depends on the listening distance.
The weighting chosen for PIR is not guaranteed in practice.

Read the book chapter of Sound Reproduction § 18.2.4 The Relationship Between Anechoic Data and Room Curves

"The real world has an infi nite variety of room shapes, sizes, and acoustical
treatments, and each one may require a different “blend” of curves from the
anechoic data set to perfect the match to a measured room curve. Doing this
in each of the infi nite number of real-world examples is obviously unproductive,
and unnecessary, because it is evident that human listeners have abilities that
surpass those of a microphone and spectrum analyzer; the room curve is not
the defi nitive answer. The message of real value here is that the right set of
anechoic data, presented in the right fashion, does a creditable job of not just
predicting room curves but, more important, permitting us to examine in detail
the temporal sequence of sounds: direct, early refl ections, and late refl ections
that listeners hear."

 

[aarons915]

What I'm referring to is that the PIR suggested that it would be a bright sounding speaker. It's revealing, but not really bright, according to every review I've read.

That looks like a good result.

I immediately thought the BMR was bright when I heard it and a few others in the review thread did actually mention the same. The ER curve in the audioholics review clearly shows the range of the BMR from 1-3k is probably responsible.

The problem with some of these internet direct companies is they have such a strong "fan base" that you will very rarely hear a negative review anywhere. In the study about sighted bias it was shown that the brand of a speaker had more influence than any other variable on perceived sound quality, the mind is powerful and is why blind tests are important in audio.

 

[Chromatischism]

I immediately thought the BMR was bright when I heard it and a few others in the review thread did actually mention the same. The ER curve in the audioholics review clearly shows the range of the BMR from 1-3k is probably responsible.

Hmm, you'd be the first I've seen. I think if the bass wasn't shelved it would be a different story.

Still, if you saw Erin's review, the response was significantly above the PIR line, by 3-4 dB.

 

[aarons915]

Hmm, you'd be the first I've seen. I think if the bass wasn't shelved it would be a different story.

Still, if you saw Erin's review, the response was significantly above the PIR line, by 3-4 dB.

They didn't sound bass shy to me at all, just an emphasized treble region and the audioholics review clearly shows a downward sloping response and then a clear peak from 1-3k (BMR range) that is more likely responsible. This is consistent with what I'm seeing with the R3, even though they are a bit laid back overall, the ER curve drops and then rises back up at the 2700Hz peak and there is a bit of brightness because of it. If you look at any Revel speaker, the ER curve is either declining or constant but never rises back up because that would basically show a directivity mismatch.

 

[edechamps]

But in both cases, PIR is not representative of each particular case.
See the graph above on the direct and diffuse distribution, it depends on the listening distance.
The weighting chosen for PIR is not guaranteed in practice.

Sure. It doesn't have to be perfect. It just has to be better than the alternatives. In the context of my PIRDI idea, the alternatives are ERDI and SPDI, which, taken individually, are arguably worse than PIRDI at estimating how we perceive reflect the directivity of the speaker in any given room. The same reasoning applies to the Olive model: one could argue that maybe feeding the actual listening test in-room response into PCA would have resulted in a better model (although that's far from a given, because it would contradict other research that says that the room has little impact in how we rank speakers), but since that wasn't done, PCA went for the next best thing and selected PIR.

(Also: I would not underestimate the accuracy of PIR. It's a common occurrence to see people in a state of shock when they measure and realize how close to PIR their own in-room curve looks like, despite what intuition would suggest. Well, except in the bass, obviously.)

 

[JIM_82]

But the spinorama like the olive notation are not tools to serve as a basis for equalization.
The rules are known and respected in terms of equalization on active monitors, flat direct and accident-free directivity.

Moreover, the subject is answered by Tool in his book.

§20.2 THE OLIVE EXPERIMENTS
...
Of great interest, of course, is which one of the several
measured curves for these loudspeakers does this correspond to, the logical
candidates being sound power (the CU model), the PIR (the in-room response
relied upon by most room-equalization schemes), or the on-axis curve?
...
From empirical evidence accumulated over many years (some of it shown
in Chapter 18), it is evident that loudspeakers with flat and smooth on-axis
frequency responses and well-behaved directivity are the ones that win listening
tests. Such loudspeakers yield smooth room curves above the transition frequency,
but they are not fl at because with typical cone and dome designs,
directivity index (DI) rises gradually with frequency, meaning that room curves
and sound power both fall gradually with frequency. Thus began the guessing
game of which “target” room curve is ideal for which circumstance (see Section
18.2.6 for a discussion of the “X” curve). Obviously, it must depend on the
directivity of the loudspeaker and the refl ective properties of the room.


So, no weighting of any anechoic measurements.

 

[Chromatischism]

When discussing the topic with us at AVS, Toole stated repeatedly that he would not equalize without "comprehensive anechoic measurements", and if not using comprehensive anechoic measurements, certainly do not EQ above the transition frequency of the room.

"Comprehensive anechoic measurements" means the data contained in a spinorama, not predicted in-room.

 

[aarons915]

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 had some more time to make some filters that closely mimic the Be line since yours were very smooth, I really wanted to give them more of a chance while adding a bit more in the 1-3k region, basically split the difference between your filter and what I was using. The Be line looks like it aims for an ER curve about 3 db below the listening window so I did the same and have been comparing to my other curve that is slightly more neutral in the LW but has a gently declining slope in the ER curve. Going back and forth I think I'm liking the Revel style ER curve now. It's not a huge difference but I notice that with the gently downward sloping curve, female vocals and the Violin can be a bit harsh sometimes, especially at louder volumes. The Revel style ER curve is just really smooth and I'm sure it's not by chance considering most of their new speakers are following the same trend.

Anyway here are the newest best filters I've found for the KEF R3 lol, subject to change next week. It would be nice if someone with the R3 would check these out and give some subjective impressions.

 

[QMuse]

If you take ON (or LW), and subtract PIR, you end up with something that's a directivity index, like ERDI and SPDI, but with different weights. Let's call it PIRDI (a term I've never seen used anywhere, I might have just invented it!) I'm speculating here, but it's quite possible that PIRDI is a better indicator of how we perceive speaker directivity compared to ERDI or SPDI, because it takes into account the total reflection from a typical room.

True, I like this idea!

My hypothesis is: since PIRDI was not in the list of curves Olive used to train the preference rating model, PCA selected PIR instead, because that's the next best thing. Maybe if PIRDI had been an option, the model would have used ON+PIRDI instead of ON+PIR. The reason why I'm making this hypothesis is because it would make Olive's model (which presently uses PIR) consistent with @Floyd Toole's conclusions (which says we don't hear the in-room response, at least not directly). Such an alternative model would make a better case for causality, as opposed to just correlation.

This has significant implications for optimizing EQ: if ON/LW+PIR is what matters, then we should take PIR smoothness into account when EQ'ing. If ON/LW+PIRDI is what matters, then we should only optimize for ON/LW when EQ'ing, because PIRDI can't be changed by EQ.

Eh.. Any directivity index, PIRDI included, is a poor choice for scoring for the same reason it cannot be changed by EQ.

Imagine this situation: you have perfectly flat ON/LW and ideally smooth PIR tilted 10dB over 20HZ-20kHz range. PIRDI is a straight line representing the difference between ON/LW and PIR. Everything perfect so far, right?

Now imagine a speaker that has +6dB peak at 1000Hz eqivalent to Q=3 on that flat ON/LW line. Same peak exists in PIR. That is far from perfect, right? But PIRDI is still that same straight line as it represents the difference between ON/LW and PIR. So, we have same PIRDI for a perfect case and the case that is far from perfect, which makes PIRDI (and any other DI) a poor choice for a candidate in the scoring system.

 

[QMuse]

"Comprehensive anechoic measurements" means the data contained in a spinorama, not predicted in-room.

You seem to fail to realise that predicted in-room response (PIR) is part of CEA-2034 standard (aka "spinorama") and is calculated from those same "comprehensive anechoic measurements".

 

[edechamps]

Imagine this situation: you have perfectly flat ON/LW and ideally smooth PIR tilted 10dB over 20HZ-20kHz range. PIRDI is a straight line representing the difference between ON/LW and PIR. Everything perfect so far, right?

Now imagine a speaker that has +6dB peak at 1000Hz eqivalent to Q=3 on that flat ON/LW line. Same peak exists in PIR. That is far from perfect, right? But PIRDI is still that same straight line as it represents the difference between ON/LW and PIR. So, we have same PIRDI for a perfect case and the case that is far from perfect, which makes PIRDI (and any other DI) a poor choice for a candidate in the scoring system.

Well, if you're saying that PIRDI alone is a poor choice for a scoring system, then I 100% agree with that. If I could only use one curve for scoring speakers, it would be ON/LW, consistent with the @Floyd Toole philosophy "if the direct sound is wrong, nothing else matters". But in practice we don't have to constrain ourselves to a single curve, so I would use both ON/LW and PIRDI.

Here's why I disagree with the conclusions you draw from your example: the directivity of the speaker you're describing is perfect. And my proposed scoring system that uses ON/LW and PIRDI will penalize the speaker, because it has a peak in ON/LW. It won't penalize it because its directivity is wrong (it's not), or because its PIR is wrong (I say we don't hear that directly), it will penalize it because its direct sound is wrong. In other words, it will penalize it for the right reasons (causality), at least according to my hypothesis.

Imagine the opposite situation. Let's say I have a similar speaker with the same peak in ON/LW, but its PIR doesn't have that peak (i.e. PIR is perfect). Is my speaker better than yours? According to your philosophy, it is, because PIR is better. According to my philosophy, my speaker is actually worse because it has a directivity error that your speaker doesn't have.

 

[QMuse]

Well, if you're saying that PIRDI alone is a poor choice for a scoring system, then I 100% agree with that. If I could only use one curve for scoring speakers, it would be ON/LW, consistent with the @Floyd Toole philosophy "if the direct sound is wrong, nothing else matters". But in practice we don't have to constrain ourselves to a single curve, so I would use both ON/LW and PIRDI.

Here's why I disagree with the conclusions you draw from your example: the directivity of the speaker you're describing is perfect. And my proposed scoring system that uses ON/LW and PIRDI will penalize the speaker, because it has a peak in ON/LW. It won't penalize it because its directivity is wrong (it's not), or because its PIR is wrong (I say we don't hear that directly), it will penalize it because its direct sound is wrong. In other words, it will penalize it for the right reasons (causality), at least according to my hypothesis.

Imagine the opposite situation. Let's say I have a similar speaker with the same peak in ON/LW, but its PIR doesn't have that peak (i.e. PIR is perfect). Is my speaker better than yours? According to your philosophy, it is, because PIR is better. According to my philosophy, my speaker is actually worse because it has a directivity error that your speaker doesn't have.

What I wanted to say is that PIRDI can't handle the situations in which both responses deivate the same way because PIRDI stays the same. For that reason you need to use criteria that both responses comprising PIRDI are smooth which then results in smooth PIRDI as well. So, criteria when both responses are smooth imply that PIRDI is smooth but when PIRDI is smooth it definitely doesn't imply that both responses comprising it are smooth. Hence PIRDI is not usable for scoring.

 

[edechamps]

What I wanted to say is that PIRDI can't handle the situations in which both responses deivate the same way because PIRDI stays the same. For that reason you need to use criteria that both responses comprising PIRDI are smooth which then results in smooth PIRDI as well. So, criteria when both responses are smooth imply that PIRDI is smooth but when PIRDI is smooth it definitely doesn't imply that both responses comprising it are smooth. Hence PIRDI is not usable for scoring.

Sure. I have never claimed that PIRDI alone is usable for scoring speakers. PIRDI alone will only tell you if the speaker has consistent directivity. It will not tell you if the direct sound is correct.

I can say the same thing about any curve, you know. For example, one can also argue that PIR is not usable for scoring because it tends to bury issues that only affect the direct sound, and some experts (notably @Floyd Toole) disagree that it directly relates to what we hear. Or that ON is not usable for scoring because it doesn't take directivity into account. Or that LW is not usable for scoring because it doesn't really reflect the direct sound one would get at any given angle. Etc.

We can do this all day. I don't believe a single curve is the best way to score speakers. But I do believe a single curve (ON or LW) is the best way to set an EQ target, because when EQing the PIRDI doesn't change.

 

[bobbooo]

The same reasoning applies to the Olive model: one could argue that maybe feeding the actual listening test in-room response into PCA would have resulted in a better model (although that's far from a given, because it would contradict other research that says that the room has little impact in how we rank speakers), but since that wasn't done, PCA went for the next best thing and selected PIR.

Section 6 of Olive's preference formula paper (part 2) tests this, with the results suggesting PIR is in fact as good if not a better predictor of preference than actual in-room response (at least when the formula is only based on these curves alone), with the correlation being 0.94 for PIR, and 0.91 for in-room response (both 1/20-octave smoothed):

 

[edechamps]

Section 6 of Olive's preference formula paper (part 2) tests this, with the results suggesting PIR is in fact as good if not a better predictor of preference than actual in-room response (at least when the formula is only based on these curves alone), with the correlation being 0.94 for PIR, and 0.91 for in-room response

Nice, I forgot about that part of the paper. This does lend credence to my idea that the Olive model selected PIR not because a good in-room response causes good sound, but because a well-behaved PIR correlates with a better-sounding speaker in general. (And from there, I'd speculate that's because a well-behaved PIR correlates with a well-behaved PIRDI, which is the actual cause the model was looking for, but couldn't use as it wasn't in the list of input curves.)

 

[Robbo99999]

Nice, I forgot about that part of the paper. This does lend credence to my idea that the Olive model selected PIR not because a good in-room response causes good sound, but because a well-behaved PIR correlates with a better-sounding speaker in general. (And from there, I'd speculate that's because a well-behaved PIR correlates with a well-behaved PIRDI, which is the actual cause the model was looking for, but couldn't use as it wasn't in the list of input curves.)

Have you thought about how ON or LW could be combined with PIRDI in a formula to represent a new scoring system? Perhaps if you've got some solid vision on how to present it in a formula then a few of us could try EQ'ing to your new scoring system to see if a higher score equates to better sound.....albeit that's just the same as EQ'ing to LW or ON? To test the solidity of the formula some more you could find two speakers that have close/identical PIRDI and then once both speakers are EQ'd to same LW then you'd expect both to sound as good as each other.....and then you'd listen to both speakers to see if you could tell a difference or preference.......given doing the stuff in this last sentence is gonna be a bit challenging because probably not many people with x2 different speakers that have the same PIRDI combined with spins from Amir and combined with being able to do blind tests - very slim chances for significant numbers of people for testing.