As someone following this discussion in the background, or at least trying to follow everything - one thing confuses me...
I see lots of talk of correcting one point or designing for one axis but in my mind a point or an axis is infinitely small so just beacuse you sit in the same seat every time you listen doesn't mean it's posible to be listening on one axis or at one point. The ears are spaced, there's natural movement of the head and you always sit slightly diferently in the chair.
So if correcting something at one point makes it worse everywhere else, surely most of the time at least one ear will be in one of those everywhere else places?
Which kind of makes it a pointless excercise? Or am I missing something?
John
I see lots of talk of correcting one point or designing for one axis but in my mind a point or an axis is infinitely small so just beacuse you sit in the same seat every time you listen doesn't mean it's posible to be listening on one axis or at one point. The ears are spaced, there's natural movement of the head and you always sit slightly diferently in the chair.
So if correcting something at one point makes it worse everywhere else, surely most of the time at least one ear will be in one of those everywhere else places?
Which kind of makes it a pointless excercise? Or am I missing something?
John
When it's said that some adjustment will only fix for one point in space, that adjustment may be a bad thing. Leaving it is likely to keep a better overall balance.
The thing to look out for as I see it, is that these 'adjustments' are often electrical, yet these issues often have a fix which is acoustic in nature.
The thing to look out for as I see it, is that these 'adjustments' are often electrical, yet these issues often have a fix which is acoustic in nature.
Right, that's how I understood it... Same for a "design axis" though right? What are the chances you happen to be on that exact axis that you've "fixed"? So fixing a design axis is likely to make things worse overall...
If you measure from a point, e.g. the listening position, then you'll get reflections and so forth included. I'd see those as unique for that point in space.
If you find, by the way, that it's coming down to that I'd want to make sure I was doing the measuring correctly and searching for the right things.
If you find, by the way, that it's coming down to that I'd want to make sure I was doing the measuring correctly and searching for the right things.
The variations in the FR from this level of change of axis are small enough to matter very little, if at all. The discussion also is not necessarily designing only to a single axis, it's about what occurs at all axes and what is the influence from the other axes as well. It does raise the issue of what the importance of the listening axis vs. the overall polar response.
For non-waveguide systems with significant baffle diffraction (waveguides minimize bafle diffraction in the first place) it may be better to not correct for it on that axis with the exception of the peak in the baffle step area due to detrimental effects in the polar response as well. That's one reason for offsetting drivers (tweeter primarily) and for applying diffraction control (felt, roundovers). You correct the problem at the source with both improved direct and polar response. I do that and optimize on the listening axis because I know that the polar response (horizontal primarily, some vertical) is also improved with both the diffraction control and subsequent optimizations.
A lot of folks still do not want to go the latter primarily due to aesthetics. To some (I would say many if not most) that takes precedence.
Dave
Last edited:
Loudspeakers tend to have two types of problems. In my experince, about 1/2 of them can be corrected by electrical means, but the other half are only made worse by electrical means. The goal of what I do, with the measurements that I take etc., is to sort out which is which and to "fix" the acoustical ones acoustically and the others with EQ. This is not as easy as it may sound.
Not easy maybe but completely logical - I come across a lot of people who use EQ as a cure-all and I've never quite been comfortable with "don't wory we'll just EQ it flat at the listening position" for the very reasons discussed here. Also minimizing the source of problems seems a much more worthy goal than trying to fix the results of problems...
Hi again,
I read (which I had read a few years back but forgotten) your study, you are referring to this one I think:
http://www.gedlee.com/downloads/AES06Gedlee_ll.pdf
I have some questions on this and some comments, if you don't mind 🙂
I am trying to replicate the FR curves given in Figure 1 and Figure 2, but I can't. For three reasons I think:
- I don't know what order and what kind of filter is used for the 2Khz high pass signal. Is it possible if you can give these?
- The levels given 2, 4 and 6db, do they correspond to the levels at 10Khz with 0 msec delay, what do they exactly correspond to?
- When doing the sum of the signals, is the high pass filtered signal delayed, or the original signal is delayed (If I delay the original signal then I can get dips at similar frequencies as Figure 2, if I delay the high pass filtered then the initial peaks and notches gets reversed.) This also dictates if the sum is minimum phase or not. If original signal is delayed sum becomes non-minimum phase, otherwise not as long as the filtered signal is smaller than the original in magnitude.
The reason I am asking these, I noticed something in the data presented that could be explained not by the group delay, but by the resulting frequency response amplitude changes. One of the things I saw interesting was that 0.6 msec delay was most often scores higher than 0.8 msec delay. And sometimes 0.2, 0.4 and 0.8 score similarly. This later set being doubles of each other, result in similar FR dips and notches, but 0.6 msec and 1msec don't follow them in the FR curves.
So I would like to reproduce the FR curves as for the exact system used in the study to take a more accurate look.
If it is easier, if you can provide the FR curves (only two are given in the paper) that would also help.
Last edited:
To give some more info on my line of thoughts, I tried to recreate the data from your paper I referred above to Excel (From the graphs of the paper). The paper gives the scores by delay and SPL, but not by amplitude (like 0db, 2db, 4db and 6db) . These amplitudes define the level of the peaks and the notches of the comb filters created by the summation of the delayed signal. The listeners were not subject to a delayed signal but to these amplitude changes of the FR also, which for the minimum phase case, these can't be separated from one another in the linear system.
So below is the amplitude vs score graph I extracted from the data on the paper. It gives the average score for all delay values for each SPL and amplitude. As can be seen, the scores are highly correlated with the amplitude of the peaks and notches in the FR amplitude, more so than the correlation of the score with delay and SPL:
And this is the overall average of all scores (SPL and delay) by amplitude, in short average of the curves above to give an overall picture:
Another thing that looks interesting to me is that, in the 1st graph, the curves for 77 and 80 db SPL are very close to each other, this can be interpretted that as the SPL is increasing, the subjects are able to hear the differences better, but this increase is reaching to its limit with the increase. Otherwise 77db and 80db should have been further apart, not that close. This can also be interpretted by the Fletcher-Munson curves, which is not something unexpected if looking from this angle.
So below is the amplitude vs score graph I extracted from the data on the paper. It gives the average score for all delay values for each SPL and amplitude. As can be seen, the scores are highly correlated with the amplitude of the peaks and notches in the FR amplitude, more so than the correlation of the score with delay and SPL:
And this is the overall average of all scores (SPL and delay) by amplitude, in short average of the curves above to give an overall picture:
Another thing that looks interesting to me is that, in the 1st graph, the curves for 77 and 80 db SPL are very close to each other, this can be interpretted that as the SPL is increasing, the subjects are able to hear the differences better, but this increase is reaching to its limit with the increase. Otherwise 77db and 80db should have been further apart, not that close. This can also be interpretted by the Fletcher-Munson curves, which is not something unexpected if looking from this angle.
These are SPL score curves for each delay, for comparison with amplitude scores:
This is the overall average score of the SPL's, not here also that the SPL scores has trend of "saturating", the difference between 77 and 80 db is not as much as others:
This is the overall average score of the SPL's, not here also that the SPL scores has trend of "saturating", the difference between 77 and 80 db is not as much as others:
These are the delay scores by SPL. Compare this to the amplitude scores graph. Amplitude scores have very steady and strong correllation with score in comparison. (Same can be said for SPL curves) Also note lower score of 0.8 msec delay in between 0.6 msec and 1.0 msec:
This is the overall average of delay scores, esiar to see 0.8msec is lower than 0.6msec:
As I wrote in my earlier message, the stronger correllation to FR amplitude peak and dips and the fact that 0.8msec scores lower than 0.6msec and 1msec delays makes me conclude the differences heard are more related to the FR amplitude non-flatness than the delayed sum signal's delay amount.
If I could get the exact FR curves of the sum signals used in the study, I could see better at what frequencies are the peaks and notches. I did generate my own curves and I can see something in those, but I don't want to speculate more since I couldn't match exactly the curves given in the paper.
This is the overall average of delay scores, esiar to see 0.8msec is lower than 0.6msec:
As I wrote in my earlier message, the stronger correllation to FR amplitude peak and dips and the fact that 0.8msec scores lower than 0.6msec and 1msec delays makes me conclude the differences heard are more related to the FR amplitude non-flatness than the delayed sum signal's delay amount.
If I could get the exact FR curves of the sum signals used in the study, I could see better at what frequencies are the peaks and notches. I did generate my own curves and I can see something in those, but I don't want to speculate more since I couldn't match exactly the curves given in the paper.
Interesting analysis Feyz 🙂
I'm not sure if the exact frequencies of the peaks and dips are important as such (perhaps it is in the critical 3-4Khz area where the ear is extra sensitive and easily irritated, but probably not elsewhere) but rather the non-flatness, and the slope of the amplitude change between each peak and dip.
I've long noticed a correlation between narrow band flatness of frequency response and perceived quality at high SPL's, but never really looked into why that might be the case.
At low SPL's a speaker that has significant narrowband fluctuations in response but has an overall neutral tonal balance (1/3 octave averaged result is balanced) can sound fine, in fact sometimes sound "more interesting" or "more visceral" at lower SPL's than a speaker with much flatter narrowband response.
However at high SPL's the speaker with significant narrowband fluctuations will sound "aggressive" and "fatiguing", at least to me, and I find that sets a ceiling on the maximum volume that I'm willing to listen to that speaker, even if distortion is not an issue.
On the other hand the speaker with a very flat narrow band response, (no sudden steep narrow band up and down slopes in the response) will tend to sound a bit more "laid back" at lower SPL's and perhaps less "impressive" than the first speaker, but at higher SPL's will sound far better than the first speaker, and in fact can be listened to much louder without any fatigue or sense of "harshness".
Whats interesting about this is that even if speaker number 2 has a modest overall tonal imbalance across the spectrum (very gradual low slope shift in response over a couple of octaves or more with no sharp discontinuities) to me it still generally sounds better than the speaker which has a balanced 1/3 octave response but narrow band peaks and dips.
So when we talk about the frequency response of a speaker I think we need to distinguish two different things - overall tonal balance (1/3rd octave averaged response) and "smoothness". (narrow band smoothness - absence of steep up and down slopes in the response over small frequency ranges, especially at high frequencies)
Although the ultimate goal is a flat response, missing that goal in these two different ways has a very different effect on the sound quality, and I think many designs in the past that might have "measured well but sound bad" actually did measure bad, but the designer was focusing more on overall tonal balance and not looking at narrow band smoothness. (Bending the 1/3rd octave response to be flat with the network, but not dealing with narrow band phenomena like high Q resonances or diffraction)
While I agree with Earl's research that diffraction does indeed cause harshness which is far more noticeable at higher SPL's than lower, particularly from about 2Khz upwards, like you I'm not convinced that "group delay" is the explanation, and I have often thought that its probably just a frequency domain phenomena.
One reason I think this is that I find that multiple side by side high Q resonances (such as often seen in poorly damped cone breakup resonances) tend to give a very similar "harshness" at high SPL as diffraction. Although not identical in character, both can result in a narrow band frequency response that is ragged like a saw blade, with very steep up and down slopes, and I believe this is the real problem.
Part of the perceptual dislike of a response like this I think comes down to amplitude modulation of any signal which contains frequency modulation, (eg vibrato) I'm sure we've all noticed at one time or another a singer on a certain song on a certain speaker applying vibrato to their voice causing an uncomfortable harshness - as their note hits the steep slope at the edge of a high Q resonance or other sharp frequency response discontinuity such as one caused by diffraction the amplitude is modulated, sometimes dramatically so. Listen to the same song on a different speaker with a smooth narrow band response in that frequency range and the harshness is gone.
I believe to completely eliminate this "harshness" and have a speaker that sounds good at high SPL you have to both eliminate high frequency (>2Khz) diffraction, (certainly doable) and also eliminate high Q resonances in the same frequency range (somewhat more difficult) if both of those are dealt with you will automatically end up with a response that has no narrow band fluctuations in response. (Although will probably still need broad band EQ to reach an overall tonal balance)
If you're trying to identify factors that correlate with sound quality preference at high SPL - try correlating both the slope of the narrow band frequency response variations (dB per 1/n octave) and also the density of the slope reversals. (slope reversals per 1/n octave) I think that will give you a pretty good correlation 🙂
Last edited:
Thanks for the nice post and sharing your experiences, it is informative for me.
The way the signal is processed in the paper, the peaks and dips start slowly after 1Khz and then gradually turn into full comb filter. So yes, there are differences in where the peak and dip is in the ear's critical hearing range depending on which delay is used, which is I suspect is one of the main reasons for the way scores have been. If I delay the main signal and add it to the HP filtered 2Khz (which is the only way I could get a match to the shape of the given FR in the paper in Figure 2) then, 0.8 msec delay gives a dip at 3Khz, 0.4msec is dipping there, low in amplitude. 0.6msec and 1msec on the other hand has a peak at 3Khz.
Note that by delaying the main signal, the end result becomes non-minimum phase. If the HP filtered signal is delayed, the end result is minimum phase. Diffraction is minimum phase, so if indeed the HP filtered signal was delayed in the study, I don't know how much the study can be related to diffraction from point of group delay.
I am not saying the following can be the only thing happening in your experience, but something that needs to be eliminated to come to conclusions: the FR amplitude shape, whether result of cone breakups or diffraction shapes the non-linear distortion profile of the driver. As an example this is often seen in distortion sweeps of metal drivers that have a high breakup peak, the frequency region that is 1/3rd of the breakup peak usually ends up having a peak at 3rd harmonic distortion component. So this may make such driver more harsh at high SPL (motor generating more distortion and then amplified by the cone breakup) vs a driver without such peak but same motor. This needs to be eliminated when making comparisons.
Agreed only with the note that baffle diffraction, other than a case of a small driver on a circular baffle and on axis, doesn't give high Q resonances. In a rectangular baffle the diffraction signal arrival is distributed in time by the different path lengths from cone to baffle edges which smooths the FR curve compared to a pure comb filter obtained by just adding delayed signal back to original.
Again typical baffle diffraction which is of a rectangular baffle doesn't give saw blade (comb filter) response, it gives a smoothed version of it. Even the size of the driver surface area smooths the diffraction.
The density of slope reversals (number of notches and peaks) in a comb filter like this study increases with the increased delay of the added signal. At 0.1msec delay you get notches at every 10Khz, at 1Khz you get notches at every 1Khz. So if we go by this, 0.8 msec delay used in the study should have given higher scores than 0.6msec delay, but didn't. That's why I am interested where exactly the peaks and notches fell for the delays used in the study.
I understand your point, the way I tried to eliminate distortion differences between different drivers as a cause in some of my testing was by using the same driver in both cases - I did some testing with some full range drivers with specific known moderately high Q cone resonances that would lead to audible harshness at high SPL if left uncorrected, and corrected each individual resonance (between 4-6 depending on driver) with a digital PEQ where gain, bandwidth and centre frequency for each resonance was very carefully adjusted based on both measured narrow band frequency response smoothness and obtaining the best early (20dB) decay in a waterfall measurement. (essentially the same thing...)
The unambiguous subjective result was that harshness at high SPL was greatly reduced, to the point where it was barely noticeable, if at all. The change was quite transformative in fact. Therefore I don't believe non-linear distortion was a contributing factor, at least in the drivers tested, as the EQ would not have made any improvement to distortions relating to cone breakup modes.
That was a point I was actually trying to make - diffraction is not a resonance, indeed, but both resonances and diffraction can cause narrow band non-flatness in the response, diffraction more so at very high frequencies, and that it may in fact be the non-flatness of frequency response that matters, not the way in which it is caused.
In my testing I found even a single 1dB 1/3rd octave peak at 8Khz was enough to add a very noticeable element of "harshness" to certain musical passages that was eliminated when the peak was carefully corrected.
Not with a midrange driver perhaps as its beaming by the time we get high enough in frequency where it (perception of harshness) matters. But a dome tweeter roughly equidistant from square edges near the top of a narrow cabinet is definitely going to have some serious issues.
A pure comb filter is a bit of an artificial test signal though, as you point out it doesn't correlate well with actual diffraction, which is smeared out more in time due to unequal distances.
The reason why I suggested that the narrowband slope of the curve might be important is because just stating the amplitude range between peaks and dips (as in your first graph) can't be enough information to describe the situation - because you can easily imagine a scenario where the peaks and dips have the same amplitude variation but are spread out much further in frequency, and this may not add a "harshness" characteristic, in fact if spread out far enough it will be a tonal imbalance over a wide frequency range. As the peaks and dips get closer (and the slopes increase) a point must be reached where it is perceived as harshness.
Although its just an intuitive surmise, I believe that harshness falls within certain boundary limits for slope and density of narrow band response variations, but theres no doubt in my mind that the harshness comes from those steep narrow band variations in frequency response, whatever their root cause.
Last edited:
If this is your point in all this then I don't have a problem with it (except that the .6 ms. versus the .8 ms differences are most likely not statistically significant). The frequency response changes would be the strongest correlation if looked at as a stand alone effect (it was not even considered as a test variable since it is simply the end result of the combined effects of the various test variables). But that completely misses the point. The correlation with the delay time and the SPL level were also statistically significant and these effects have never been reported before and are for the most part completely ignored.
I don't remember much about the details of the experimental setup, and I am not too interested in looking them up since nothing in what you are saying would change our conclusions at all. I do remebre that the HP signal was delayed, but it could have been subtracted instead of added I don;t remember (diffraction usually causes a sign change to the diffracted signal.) The filter slope I would suspect as very low order, probably first or second as I can see no physical rational for anything higher. Beyond that duplicating the FR curves should be quite trivial.
But that is completely semantic in that the group delay causes the FR variations, they are one and the same situation. If you remove the group delay the FR variations go away. Hence the FR variations and the group delay are completely correlated and so if the audibility is correlated to the FR variations then it is by inference correlated to the group delay. You are trying to seperate two things that are inseperable.
That is not always true. My main interest was the audibility of Higher Order Modes (HOM) in a waveguide and here they ARE NOT smeared out in time as they would be for "rectangular" baffle diffraction. A single HP time delayed signal is quite accurate for this case. Let's not over simplify everything by making global assumptions that don't hold in many cases.
This was a preliminary study (just try and do valid double blind subjective studies, they are very difficult and time consuming, even simple ones) and as such very simple techniques and signals were used. The hypotheiss that group delayed signals could have a "non-linear" audiblity that depended on "linear" parameters was proven. You cannot extrapolate the results much further than that. But the study does imply that there is more to audibility than is usually assumed. Is more work required? Sure! Is this going to happen? Not likely!
These results fundamentally changed the way I thought about perception and in particular the perception of "nonlinear" distortion. That a linear perturbation could be perceived as a nonlinear effect was a real paradygm shift for me. I got what I needed out of this study and moved on. I'm more interested in other dynamics effects at this time.
I mean no disrespect, but above two, to me you are conflicting. And in the paper it is represented as mostly group delay is being studied, and as if it is a non-minimum phase group delay result. But it is not. In the paper, and also your reference to the paper there is a strong emphasis on group-delay and no mention of amplitude response non-flatness. So if min-phase, then most people normally talk about FR amplitude, not group delay.
My main objection is, it is mixing terms and concepts and not actually revieling anything new. In the end it shows that FR amplitude non-flatness is audiable.
"That a linear perturbation could be perceived as a nonlinear effect was a real paradygm shift for me. I got what I needed out of this study and moved on. I'm more interested in other dynamics effects at this time. "
Well, the increase in perception is starting to converge to a value from the 4 different levels given, as I depicted above.
So this can be just easily the result of loudness curves. We all know to hear more detail sometimes you need to turn up the volume, so that the highs and lows get to ear's level that are more sensitive to them.
I think that you are mising the point - we studied; group delay (time in ms); level of that delay (in dB), and absolute SPL (in dBSPL). That there are resulting frequency response varaitions with these parameters was a given and noted in the paper. But the frequency response varaitions and the studied variables are not independent, they are completely correlated.
What you are implying is what was shown. Show me where this was proven before.
Do we? How did we know that? Where was this shown? Such a statement is completely contrary to linear systems theory which says that the response to a linear system is independent of the amplitude. So what you claim "we all knew" is not very intuitive and quite honestly not at all obvious.
The simple statement - "I knew that!" (or better yet "We ALL knew that.") is such a poor argument. It's so easy to say in hindsight, but unless you can show that you stated this somewhere prior to the paper then its simply your word unsubstantiated.
Simon implied that it was all just frequency response - OK, prove that with a valid subjective test. I don't believe that the results will be the same as what we got, but at any rate since there are no results, only conjecture, its a moot point.
What you are implying is what was shown. Show me where this was proven before.
Do we? How did we know that? Where was this shown? Such a statement is completely contrary to linear systems theory which says that the response to a linear system is independent of the amplitude. So what you claim "we all knew" is not very intuitive and quite honestly not at all obvious.
The simple statement - "I knew that!" (or better yet "We ALL knew that.") is such a poor argument. It's so easy to say in hindsight, but unless you can show that you stated this somewhere prior to the paper then its simply your word unsubstantiated.
Simon implied that it was all just frequency response - OK, prove that with a valid subjective test. I don't believe that the results will be the same as what we got, but at any rate since there are no results, only conjecture, its a moot point.
Last edited:
- Status
- Not open for further replies.