Sony SS-CS5 3-way Speaker Review

[QMuse]

0
True, If I stop the raw PIR before ~1000Hz, the two adjustments look almost equally different, with yours taking a slight edge:
View attachment 65329

However, If I also include the ~1700Hz-3200Hz region which is mostly a straight line, your slope is very close:
View attachment 65331

There is also the difference between lowering peaks and raising dips, BYRTT's raises 3kHz by 3dB (2x the wattage), and I don't think this speaker will like that in terms of distortion and compression.

What is interesting here is that slope isn't taken into account with scoring so they scored very closely but it definitely affected the way how @amirm perceived them.

 

[MZKM]

What is interesting here is that slope isn't taken into account with scoring so they scored very closely but it definitely affected the way how @amirm perceived them.

The formula actually does favor more tilt. It just usually isn't much of an issue, but this is why ultra wide directivity speakers may be underscored, as they weren't used when creating the formula.

 

[andreasmaaan]

The formula actually does favor more tilt. It just usually isn't much of an issue, but this is why ultra wide directivity speakers may be underscored, as they weren't used when creating the formula.

Are you sure that we know that? As I understand it, we don't know how most of the speakers in the larger sample measured.

 

[QMuse]

The formula actually does favor more tilt.

I didn't know that. We may speculate that formula would do better if more weight is put into the tilt.

 

[MZKM]

Are you sure that we know that? As I understand it, we don't know how most of the speakers in the larger sample measured.

Well, we don't know for sure, but the formula does favor a more tilted PIR curve. Instead of having the curve normalized to 0 slope or referenced to the target slope and then analyzed, it calculates the correlation coefficient, which looks for dependency, in this case between frequency and SPL, a mostly flat on-axis has no dependency, but a PIR curve that decreases in SPL the higher up you go does, and the more it decreases, the higher the dependency.

 

[andreasmaaan]

Well, we don't know for sure, but the formula does favor a more tilted PIR curve. Instead of having the curve normalized to 0 slope or referenced to the target slope and then analyzed, it calculates the correlation coefficient, which looks for dependency, in this case between frequency and SPL, a mostly flat on-axis has no dependency, but a PIR curve that decreases in SPL the higher up you go does, and the more it decreases, the higher the dependency.

Correct me if I'm wrong, but IIRC this PIR target slope tells us only about the PIR slopes of the speakers in the sample that were most preferred (since it was only the most-preferred speakers whose PIRs were used to calculate the target slope).

It doesn't tell us anything about the PIR slopes of the non-preferred speakers (i.e. the majority of the speakers) in the sample.

EDIT: my apologies. I had misread your earlier post. You said ultra-wide directivity speakers "weren't used when creating the formula". This is obviously correct (or tends to be so).

I had misread your earlier post as saying that ultra-wide direcitivity speakers were not included in the sample.

They may well have been in the sample, but if they were not among the most preferred speakers, their PIRs would not have been used to calculate the target slope.

 

[QMuse]

.. it calculates the correlation coefficient, which looks for dependency, in this case between frequency and SPL, a mostly flat on-axis has no dependency, but a PIR curve that decreases in SPL the higher up you go does, and the more it decreases, the higher the dependency.

I'm not sure I follow. Correlation coefficient measures deviation from linearity. 2 PIR curves with different slopes can have the same correlation coefficient, so how does it take slope into account?

 

[MZKM]

I'm not sure I follow. Correlation coefficient measures deviation from linearity. 2 PIR curves with different slopes can have the same correlation coefficient, so how does it take slope into account?

It looks for deviation, but sees if it can be correlated with another parameter, in this case the frequency.

Besides @edechamps helping me realize this, I also asked on Reddit. The deviations are still there, but a slope is a linear relationship between the axes, so that is a high correlation between them.

 

[QMuse]

It looks for deviation, but sees if it can be correlated with another parameter, in this case the frequency.

Besides @edechamps helping me realize this, I also asked on Reddit. The deviations are still there, but a slope is a linear relationship between the axes, so that is a high correlation between them.

It is not how it works. Have a look at this graph, blue dots are samples and red line was calculated using least squares method.

Correlation coefficient is a measure how blue dots correlate with red line. As more of them were closer to red line correlation coefficient would tend to be closer to 1.

Now imagine blue dots are rotated around the black point in the middle for some angle, say 20 deg clock-wise and that their (x,y) corrdinates are recalculated acccordingly. using least square method you would see that red line would rotate with them changing slope. But if you would recalculate correlation coefficient it would remain the same as distance from blue dots to the red line hasn't change. That means that correlation coefficient is not related to the slope of the red line but to the "average" distance of blue dots from the red line.

 

[MZKM]

It is not how it works. Have a look at this graph, blue dots are samples and red line was calculated using least squares method.

Correlation coefficient is a measure how blue dots correlate with red line. As more of them were closer to red line correlation coefficient would tend to be closer to 1.

View attachment 65357

Now imagine blue dots are rotated around the black point in the middle for some angle, say 20 deg clock-wise and that their (x,y) corrdinates are recalculated acccordingly. using least square method you would see that red line would rotate with them changing slope. But if you would recalculate correlation coefficient it would remain the same as distance from blue dots to the red line hasn't change. That means that correlation coefficient is not related to the slope of the red line but to the "average" distance of blue dots from the red line.

That’s what I assumed, but rotating the data while mainting distance from the linear regression did in fact change the score.

 

[andreasmaaan]

That’s what I assumed, but rotating the data while mainting distance from the linear regression did in fact change.

View attachment 65365

Just to clarify, the discussion we are having here relates to SM_PIR, but not NBD_PIR or NBD_ON. Is that correct?

 

[MZKM]

Just to clarify, the discussion we are having here relates to SM_PIR, but not NBD_PIR or NBD_ON. Is that correct?

Correct.

 

[andreasmaaan]

Correct.

Thanks

And to further clarify: Assuming that all data points are equally-spaced from the PIR slope (and all else is equal), a speaker with a steeper PIR slope will always receive a better rating for SM_PIR. Correct?

And furthermore, there is no effective limit to this, i.e. a hypothetical speaker with a PIR slope approaching ∞ dB/decade would (all else equal) achieve a superior rating for SM_PIR than a speaker with a shallower PIR slope (once again assuming all data points are equally-spaced from each speaker's respective PIR slope). Also correct?

 

[QMuse]

That’s what I assumed, but rotating the data while mainting distance from the linear regression did in fact change.

View attachment 65365

Uh, you are right - least square sum remains the same but correlation indeed changes with the slope. It's been long since I was working with that stuff - obviously too long.

 

[QMuse]

Thanks

And to further clarify: Assuming that all data points are equally-spaced from the PIR slope (and all else is equal), a speaker with a steeper PIR slope will always receive a better rating for SM_PIR. Correct?

Correct.

And furthermore, there is no effective limit to this, i.e. a hypothetical speaker with a PIR slope approaching ∞ dB/decade would (all else equal) achieve a superior rating for SM_PIR than a speaker with a shallower PIR slope (once again assuming all data points are equally-spaced from each speaker's respective PIR slope). Also correct?

Probably also correct, as even in that case correlation wouldn't actually change much, it will move slowly to 1.

 

[MZKM]

Thanks

And to further clarify: Assuming that all data points are equally-spaced from the PIR slope (and all else is equal), a speaker with a steeper PIR slope will always receive a better rating for SM_PIR. Correct?

And furthermore, there is no effective limit to this, i.e. a hypothetical speaker with a PIR slope approaching ∞ dB/decade would (all else equal) achieve a superior rating for SM_PIR than a speaker with a shallower PIR slope (once again assuming all data points are equally-spaced from each speaker's respective PIR slope). Also correct?

Here is the Sony:

Normal (scored 0.77):

Mostly Normalized (scored 0):

Steepened (scored 0.99):

The offset of the NBD score of the PIR being worse doesn't seem to be enough to counteract this.

I really wish Olive used the normalized PIR and only did NBD on it, that way the difference of slope wouldn't have such a drastic difference.

 

[andreasmaaan]

Correct.

Probably also correct, as even in that case correlation wouldn't actually change much, it will move slowly to 1.

Thanks also

Given that correlation coefficient biases SM in favour of steeper slope, why do we speculate that Olive's model uses it (as opposed to a formula that is independent of it)?

 

[QMuse]

Thanks also

Given that correlation coefficient biases SM in favour of steeper slope, why do we speculate that Olive's model uses it (as opposed to a formula that is independent of slope gradient)?

Not sure, it was just a thought.

@MZKM Can you plz post score components of @BYRTT and my filtered response?

Are this components and weighting you're using?

• Narrow-band deviation of the on-axis curve(NBD_ON): 31.5%
• Bass extension(LFX): 30.5%
• Narrow-band deviation of the predicted in-room response(NBD_PIR): 20.5%
• Smoothness of the PIR(SM_PIR): 17.5%

 

[MZKM]

Not sure, it was just a thought.

@MZKM Can you plz post score components of @BYRTT and my filtered response?

Are this components and weighting you're using?

• Narrow-band deviation of the on-axis curve(NBD_ON): 31.5%
• Bass extension(LFX): 30.5%
• Narrow-band deviation of the predicted in-room response(NBD_PIR): 20.5%
• Smoothness of the PIR(SM_PIR): 17.5%

It's the same weighting as the formula, so yes; I posted the radar charts so you can see the differences.

 

[andreasmaaan]

Final question for @MZKM, @QMuse et al.

It seems also the case that correlation coefficient would be biased in favour of a steeply upward-sloping SM_PIR slope - is this correct?