Selah Audio RC3R 3-way Speaker Review

[bobbooo]

That’s what I initially thought, as the ideal slope is part of the calculation, but when I change the target slope the final result is the same (multiplying and diving by the same thing will cancel out, and slope is part of the numerator and denominator). However, we should wait until we get a speaker with an in-room response that is very smooth but not tilting down to be extra sure.

This is one of the reasons I suggested in your Master Preference Rating thread to bring back the full calculation spreadsheets for all the speakers and putting them in one place, so anyone can easily go over them and maybe solve puzzles like this and spot potential errors. It's always best to be 100% transparent with data and calculations when doing scientific measurements and analysis.

I've just done a quick test with some dummy data which I believe confirms that SM is dependent on spectral tilt. I filled in the data with a linear series, creating a perfect linear scatter graph, and therefore an R^2 of 1. I then changed the data to another perfectly linear series also with R^2 of 1, but with a slope visually closer to the target slope, and this second set of data produced a higher SM score. So two sets of data, each perfectly linear (and so intuitively equally 'smooth') produced different SM values that depend on their spectral tilt. Could you (and anyone else) replicate this to corroborate?

 

[dshreter]

This is one of the reasons I suggested in your Master Preference Rating thread to bring back the full calculation spreadsheets for all the speakers and putting them in one place, so anyone can easily go over them and maybe solve puzzles like this and spot potential errors. It's always best to be 100% transparent with data and calculations when doing scientific measurements and analysis.

I've just done a quick test with some dummy data which I believe confirms that SM is dependent on spectral tilt. I filled in the data with a linear series, creating a perfect linear scatter graph, and therefore an R^2 of 1. I then changed the data to another perfectly linear series also with R^2 of 1, but with a slope visually closer to the target slope, and this second set of data produced a higher SM score. So two sets of data, each perfectly linear (and so intuitively equally 'smooth') produced different SM values that depend on their spectral tilt. Could you (and anyone else) replicate this to corroborate?

I think for a preference score, incorporating the spectral tilt makes sense. To build a speaker that the most people will like, it needs a good spectral balance when plugged straight into an ordinary room.

That’s somewhat different from a measure of performance though, and arguably adjustments of spectral tilt are a virtual necessity unless you are going to build a room around a specific speaker. If basic EQ can “fix” a speaker, then it isn’t really broken. The omnipresent dip switches on studio monitors and integrated EQ systems are strong evidence of that.

Most people here are in the business of buying speakers and not building them, so whether a product has best seller potential isn’t the same as if it would fit well in your own system and room.

 

[Hiten]

The measured Kali and Revel are also 3 way speakers.

ahh ! yes indeed. Will read those reviews. I think (quick search) no 2 way passive is reviewed. So was little bit more interested in passive 2 way and passive 3 way low fq. distortion comparison(At same volume level. I understand lots of things affect speaker design but was curious none the less. Atleast give vague idea.
Best regards.
UPDATE :
OK. I searched KEF LS50 is reviewed with distortion figures. Would be nice to have selah audio distortion measurements if it is not too much trouble.

 

[MZKM]

This is one of the reasons I suggested in your Master Preference Rating thread to bring back the full calculation spreadsheets for all the speakers and putting them in one place, so anyone can easily go over them and maybe solve puzzles like this and spot potential errors. It's always best to be 100% transparent with data and calculations when doing scientific measurements and analysis.

I've just done a quick test with some dummy data which I believe confirms that SM is dependent on spectral tilt. I filled in the data with a linear series, creating a perfect linear scatter graph, and therefore an R^2 of 1. I then changed the data to another perfectly linear series also with R^2 of 1, but with a slope visually closer to the target slope, and this second set of data produced a higher SM score. So two sets of data, each perfectly linear (and so intuitively equally 'smooth') produced different SM values that depend on their spectral tilt. Could you (and anyone else) replicate this to corroborate?

The SM score is R^2, so how could you do better than a 1?

How did you change the slope? The formula uses the abs difference of the actual slope and the target slope (-1.75 for PIR), for Cell D1 it’s:
=if(isblank(A1)=FALSE,abs(-1.75-($F$16))*(ln(A1))+$F$17,)
You can change -1.75 to whatever target slope you want and the final calculation doesn’t change, at least for me. You can also alter the data points in a consistent manner (I took actual vs predicted and increased/decreased the actual points by 1/2 the SPL difference from predicted), and the final calculation also doesn’t change for me.

 

[edechamps]

If basic EQ can “fix” a speaker, then it isn’t really broken.

Not sure I agree. That might be true for obsessed audiophiles, but when recommending speakers to a "normal" person, you can't really expect them to use EQ to fix the speaker. They're just going to plug it in and forget about it.

Also, if it's possible to fix the speaker with EQ, but the manufacturer decided not to do it, then IMHO the manufacturer should be penalized for ignoring a clear improvement opportunity. (Note that this means passive speakers will likely be penalized more often than active ones. I'm fine with that - it is a true reflection of the shortcomings of passive speakers.)

 

[Mashcky]

It's most likely not resonances around 800-900Hz, but simply vertical cancellations caused by the crossover from the woofer to the midrange dome.
Followed by a strong vertical expansion in the radiation in the range 1000-2000Hz - see "Vertical Directivity Normalized" in Post#4.

The horizontal frequency response measurements show no signs of resonance in the 700-1000Hz range - see "Horizontal Directivity" in Post#4.
The ScanSpeak 18M shows a suspension resonance in the range 800-1000Hz, but in this case it is suppressed by the crossover.

From a design standpoint, what is the solution to the hump that turns up in the directivity index but not in the horizontal plot? Is this an issue of time alignment causing the cancellation? I've having trouble understanding the pretty great looking initial plot with the messy directivity index.

 

[thewas]

Directivity depends on driver size, front baffle dimensions and shape (also waveguides) and working frequency of each chassis, so by changing appropriately those a smoother horizontal directivity can be achieved. Vertical directivity discontinuities depend also additionally on the driver distance and crossover slopes,

 

[ctrl]

From a design standpoint, what is the solution to the hump that turns up in the directivity index but not in the horizontal plot? Is this an issue of time alignment causing the cancellation? I've having trouble understanding the pretty great looking initial plot with the messy directivity index.

The difference between the plots is the scaling of the y-axis.

The vertical cancellations occur because the radiation of the loudspeaker chassis is not coincident. This means they do not radiate from the same point, at the same time.

This leads to the formation of radiation lobes around the crossover frequency in vertical direction.

The problem with this loudspeaker is not so much the hump in the early reflection index, which corresponds to a withdrawal of sound power (which is usually less of a problem), but the subsequent increase of the same.

The midrange dome tweeter in combination with the baffle radiates very wide in the range around 1-2,3kHz. The combination with the ribbon tweeter, which bundles strongly in vertical direction, makes things even more difficult.

This can be seen very well in the normalized angular frequency response. This constellation often sounds a bit harsh and aggressive at higher sound pressure levels. To counteract this, the sound pressure in the upper treble range was slightly increased on axis (this is my guess).

 

[bobbooo]

The SM score is R^2, so how could you do better than a 1?

How did you change the slope? The formula uses the abs difference of the actual slope and the target slope (-1.75 for PIR), for Cell D1 it’s:
=if(isblank(A1)=FALSE,abs(-1.75-($F$16))*(ln(A1))+$F$17,)
You can change -1.75 to whatever target slope you want and the final calculation doesn’t change, at least for me. You can also alter the data points in a consistent manner (I took actual vs predicted and increased/decreased the actual points by 1/2 the SPL difference from predicted), and the final calculation also doesn’t change for me.

Ah, I must have used an early spreadsheet (maybe one I tinkered with). I've just tried again with your latest master file and there is indeed no change in the SM value with spectral tilt (either target or measured). I concluded the latter by entering in perfectly smooth dummy data, one with all 'measured' data following a slight downward tilt (except for on-axis which I made ideally flat), and another with all flat (except a bright, upward-tilted on-axis). The latter results in a bright PIR, and the former a slightly downward-sloping PIR (closer to ideal). Here are the spreadsheets:

Perfectly Smooth w/ Flat On-Axis Response
Perfectly Smooth w/ Bright On-Axis Response

So the mystery remains why Olive included the SL variable in the SM formula, if neither the target nor the measured spectral tilt (slope) actually make a difference to the SM value. I think there must be a reason for its inclusion, so I still suspect there’s something wrong in the translation from Olive's description of the SM variable in his paper to the spreadsheet calculation. I really think it would be best to post all the data and spreadsheets in one easy to view place to get as many eyes on this problem as possible so this can be solved, as the scores might be inaccurate right now.

If you look at the final preference ratings on my dummy data spreadsheets though, the ‘speaker’ with the flat on-axis and slightly downward-sloping PIR (closer to ideal) does have a higher score than the bright ‘speaker’, which is as expected. So the other variables are at least accounting for spectral tilt, even if SM isn’t. They may only be partially accounting for it though, if the SM calculation is not entirely correct, and should actually show a dependency on the SL variable.

I've also spotted an error on the LFX tab of the master preference rating spreadsheet - the formula for cell B4 should be '=LOG10($B$3)', instead of '=LOG10($D$3)'. Could you correct that, and all the scores on your preference ratings charts? As they're incorrectly underestimating the bass extension (and so preference rating) of the speakers if based on that master spreadsheet.

 

[MZKM]

I've also spotted an error on the LFX tab of the master preference rating spreadsheet - the formula for cell B4 should be '=LOG10($B$3)', instead of '=LOG10($D$3)'. Could you correct that, and all the scores on your preference ratings charts? As they're incorrectly underestimating the bass extension (and so preference rating) of the speakers if based on that master spreadsheet.

It's not an error, just poor formatting on my end, B3 is the closest Hz lower than the -6dB point whereas D3 is the closest Hz higher, and I believe we agreed on closest one higher.

EDIT: Whoops, looks like you stated you wanted closest Hz less than the -6dB point, boosting the LFX score a bit rather than lowering it a bit (I personally still feel the closest Hz regardless would give more accurate results, but the impact is going to be small either way). The Kali would go from 5.0 to 5.1 if using only one decimal place.

 

[bobbooo]

It's not an error, just poor formatting on my end, B3 is the closest Hz lower than the -6dB point whereas D3 is the closest Hz higher, and I believe we agreed on closest one higher.

We agreed on closest Hz lower here, for the reasons I outlined in this post. I'm pretty sure you did change it at that point, but maybe it's somehow reverted back. You must have a lot of spreadsheets to deal with so I wouldn't be surprised if a mix up happened at some point!

 

[MZKM]

I edited that comment; the data should now be updated (allow for 5min auto-update intervals).

 

[amirm]

So what score do you get if on-axis is flat as is on-axis with a slight slope down? And say, extension goes to 30 Hz?

 

[bobbooo]

I edited that comment; the data should now be updated (allow for 5min auto-update intervals).

Brilliant, thanks Looks like the alternative score using NBD_LW hasn't been updated though in the 'Price : Performance LW w/ sub' chart.

 

[bobbooo]

So what score do you get if on-axis is flat as is on-axis with a slight slope down? And say, extension goes to 30 Hz?

I'm not sure exactly what you mean - do you mean flat (no tilt) on-axis and slight slope down for PIR? Or the change in score between no tilt on-axis and slightly sloping down on-axis?

 

[amirm]

I'm not sure exactly what you mean - do you mean flat (no tilt) on-axis and slight slope down for PIR?

Correct. Trying to see what a perfect speaker would get.

 

[MZKM]

Brilliant, thanks Looks like the alternative score using NBD_LW hasn't been updated though in the 'Price : Performance LW w/ sub' chart.

The NBD score is from 100Hz-12kHz.

 

[bobbooo]

The NBD score is from 100Hz-12kHz.

Ah yes, forgot it was the score ignoring LFX.

 

[bobbooo]

Correct. Trying to see what a perfect speaker would get.

Well a theoretically 100% perfect speaker should get a perfect 10 score of course, but as the calculation spreadsheet stands that only looks like it’s possible if the -6 dB bass extension point were even lower than the ‘ideal’ value of 14.5 Hz we calculated previously. That’s because the ideal target slope of the PIR according to Olive’s paper is -1.75, but using that, even with a perfectly smooth slope, the NBD_PIR variable is not the ‘ideal’ value of 0 in this case, as it considers the slope itself as a ‘deviation’. So by my rough calculations using dummy data, with the assumptions that the on-axis and LW responses are the same, and the ER, SP and PIR responses are the same, with a PIR slope of -1.75, and an LFX point of 14.5 Hz, I get a ‘perfect’ score of ~9.5. If you relax the -6 dB LFX point to 30 Hz, the perfect speaker with that bass extension would score ~8.2. Here’s the dummy data and calculation for that:

Perfect Speaker(?) with LFX at 30 Hz

Taking the NBD_PIR value of 0.15 from this ‘speaker’ with a perfectly smooth PIR with slope of -1.75, and rearranging Olive’s preference formula to solve for the LFX value needed for a score of 10, I arrive at an LFX value of 0.75 and therefore ideal -6 dB frequency x_SP = 10^LFX = 10^0.75 ≈ 5.6 Hz, which seems too low.

Now this is all assuming the ‘smoothness’ (SM variable) in the spreadsheet is correct, which I suspect it might not be, as has been noted, changing the target slope from -1.75 does nothing to the SM value, nor does changing the slope of the ‘measured’ PIR (as I showed in a previous post), even though these two slopes are included in Olive’s definition of SM and the spreadsheet calculation. So I think something has been lost or misinterpreted when translating Olive’s SM definition to the spreadsheet, which wouldn’t surprise me as his definition is not exactly clear.

 

[amirm]

Yeh, intuitively I still have my doubts about the scores.