I poked around a little to try and find the smoking gun - a plot of the Butterworth 3rd order crossover with tweeter above woofer and drivers connected with the same polarity. I found only one link that seemed to match all of these:
Off-Axis Response | Beat Stamm
You have to get down about 2/3 to the bottom before he says:
Several things caught my eyes right away:
1. I can't get rid of the cancellations, no matter how steep a filter slope I choose (although the filter slopes > 24 dB/oct look intriguing).
2. Odd ordered filters (6 dB/oct and 18 dB/oct) introduce lobing errors, i.e. the cancellations are asymmetric (with respect to the 0° axis), the sound that makes it to your ears appears bundled in what loosely looks like an earlobe, and this lobe is tilted downwards (reversing the tweeter polarity tilts this upwards).
3. Odd ordered filters introduce not only dips, but also peaks in the off axis response. This may actually be worse than dips[4].
1. I can't get rid of the cancellations, no matter how steep a filter slope I choose (although the filter slopes > 24 dB/oct look intriguing).
2. Odd ordered filters (6 dB/oct and 18 dB/oct) introduce lobing errors, i.e. the cancellations are asymmetric (with respect to the 0° axis), the sound that makes it to your ears appears bundled in what loosely looks like an earlobe, and this lobe is tilted downwards (reversing the tweeter polarity tilts this upwards).
3. Odd ordered filters introduce not only dips, but also peaks in the off axis response. This may actually be worse than dips[4].
He defines his speaker model earlier in the page. He shows a diagram of tweeter above woofer, although its not explicitly mentioned in the text. So, judging by all the other anecdotal evidence, plus this page, I'd have to say my sims are correct.
I did think of one thing that might explain your observation to the contrary - some drivers do not have the same "internal" polarity e.g. meaning positive voltage on "red" terminal causes diaphragm to move outwards. Perhaps, even though you have connected positive to "red" for each driver, one of them has the opposite behavior and is therefore "polarity inverted". I have heard of this kind of thing before... Just a guess, though.
-Charlie
JBL used the "opposite" way to designate polarity :
Black is + (forward cone motion)
Red is -
The graph I was thinking about was in an article in Audio, Dec-89, "Measuring Acoustic Phase" by Don Davis. It plots real and imaginary components of a speaker's impulse response as a 3-D spiral along z-axis where z = time and x,y are real and imaginary. It is not the same type of analysis as just a traveling wave.
A phase shift of 360 deg is not the same as 0 deg, except in a mathmatical sense where the trigonometric values repeat 360 = 0. For a real-world application, phase is not used to describe true nature, but instead group delay is relevant. Group delay is rate of change of phase with respect to frequency, where phase is in radians (not deg). Radians do not reset back to 0 when 360 deg is reached, but continue on.
I continue to emphasize this point because assuming 360 = 0 leads to the erroneous conclusion that a crossover toplogy where drivers have same absolute polarity is somehow desirable and will automatically "sound better". The configuration with the least amount of group delay is the correct choice, regardless of driver relative polarity.
The references I have cited in JAES and Audio are piling up; isn't it time you spend an afternoon in the library exploring all of this original analysis?
Last edited:
Numerous studies have been done to quantify the audibility of group delay. The human ear/brain combination is most sensitive to group delay in the range of 1 to 2 khz if memory serves correctly. And if memory serves correctly, group delay of more than a few milliseconds is about the limit of audibility for most people. At 1khz, the phase rotation of a fourth order Linkwitz-Riley crossover (one of the most popular crossovers on the planet) is about 1 millisecond. Actual group delay is below the threshold of audibility. Now because you've decided to pick up some articles and read (you should be commended for that, BTW
) does not mean others here have not done so and in some cases - done a whole lot more. Some of the finest loudspeakers ever produced use these types of crossovers - the Revel Ultima comes to mind as a prime example of a current state of the art design using a fourth order crossover very successfully (it might not be a full 4th order electrical in all instances - you'd be best served asking Kevin Voecks )So, to sum up, yes - group delay is something we should all be concerned about, especially with high order crossovers and ellipticals. But for 1st through 4th order crossovers, it's essentially a non issue.
Just getting started? Get your feet wet a few days ago and now we're ready to educate the masses?
If you actually bothered to look at a few real group delay plots you'd find that group delay is confined to a small portion of the reproduced spectrum about a given crossover. There's no audible group delay across several octaves and multiple crossovers. You clearly have a lot more reading to do. Hopefully, you will add a little bit of experimentation to the reading as nothing beats direct, first hand experience.
http://www.trueaudio.com/post_010.htm
There's an old saying - "a little information can be a dangerous thing". I applaud your enthusiasm but you have a long way to go on this subject. Reading a few AES papers will not make you an electronics expert. If you have "golden ears" or a "tin foil hat wrapped cranium", the following article will be irrelevant. If that's not the case, I suggest you read it or the groundbreaking work of Blauert/Laws on the subject of group delay audibility.
I already posted the link to the above article to Charlie a few posts back, so I was aware of it. (#38)
I began research in the late '70s early '80s by spending several weeks in the engineering library going through back issues of JAES, Wireless World, etc, checking countless cross-links as I discovered them. I have also since tried to keep up with all new relevant developments, most notably the Linkwitz-Riley analyses. It seems to me that many of the newer generation have overlooked most of the pioneering research and accept newer analysis at face value.
I have indeed done the experiments - tried every implementation of xover order 1st thru 4th with every configuration for a full 4-way system also with ability to quickly alter time-alignment. Did measurements and listened (listened first , measured later).
In order for these types of experiments to be meaningful, there has to be a way to make changes, within a matter of a few minutes, so accurate comparison can be made, and that is exactly what I did. Experimentation went on for several years and is still ongoing.
All of the issues I have mentioned here in various posts deserve carefull consideration by any serious audiophile.
You are mistaken in your observation about group delay. My experiments proved otherwise.
I began research in the late '70s early '80s by spending several weeks in the engineering library going through back issues of JAES, Wireless World, etc, checking countless cross-links as I discovered them. I have also since tried to keep up with all new relevant developments, most notably the Linkwitz-Riley analyses. It seems to me that many of the newer generation have overlooked most of the pioneering research and accept newer analysis at face value.
I have indeed done the experiments - tried every implementation of xover order 1st thru 4th with every configuration for a full 4-way system also with ability to quickly alter time-alignment. Did measurements and listened (listened first , measured later).
In order for these types of experiments to be meaningful, there has to be a way to make changes, within a matter of a few minutes, so accurate comparison can be made, and that is exactly what I did. Experimentation went on for several years and is still ongoing.
All of the issues I have mentioned here in various posts deserve carefull consideration by any serious audiophile.
You are mistaken in your observation about group delay. My experiments proved otherwise.
Last edited:
If you're experience or "findings" are vastly different from those of Blauert and Laws - I'd urge you to approach the AES and publish details of your experimentation and data. I believe an open mind is the essential pathway to enlightenment. So I will not conclude that you're a fruitloop until I've seen data that supports your strongly held view and disputes that of other published individuals. I will however remain skeptical until that incontrovertible data is made available.
You may find these simulations interesting (or not):
Above image, 4-way, 3rd BUT, driver polarity alternating.
Above image, 4-way, 3rd BUT, drivers same polarity.
Group delay at ~ 20 Hz is > 10 ms.
Any thoughts?
An externally hosted image should be here but it was not working when we last tested it.
Above image, 4-way, 3rd BUT, driver polarity alternating.
An externally hosted image should be here but it was not working when we last tested it.
Above image, 4-way, 3rd BUT, drivers same polarity.
Group delay at ~ 20 Hz is > 10 ms.
Any thoughts?
Not. When I said "findings", I was inferring something on the order of a double blind study that would actually have the credibility needed to refute the findings of individuals like Blauert who went through that kind of trouble. If you think you can convince me or the average person on this forum to believe you over Blauert with a simulation - no less!!!! - you're headed for fruitloop territory. Sorry - but that's the reality.
Beyond that - the data you posted lists group delay that is below the published threshold of audibility. As for group delay at 20 hz - that's a joke. Most people can only feel 20hz let alone distinguish it from 40 hz. To supplement your studies, you should include published findings on the audibility of bass doubling at the lowest fundamentals. Most people have trouble telling the difference between 20 hz and 40 hz let alone sensing something on the order of a 70 degree phase shift at that frequency!
Wasn't offering the simulations as any kind of proof of audible artifacts, only to show group delay for this particular configuration. You will have to listen for yourself to decide about audibility. Or you can choose to take your favorite analysis at face value and design your system accordingly.
All kinds of studies have been published to prove or disprove this or that theory. My goal is to point out previous valid studies that I have first-hand experience with that may have been overlooked today because "most people" have been conditioned to believe it is not important.
There is no way to convince anyone about psychoacoustic phenomenon because every person experiences it differently. All one can do is explain the physics and realize that there is a high probability some facet might be audible. Just because "most people" agree on a particular topic in such a subjective area does not mean all other observations should be abandoned. Stop picking on those of us in the minority.
All kinds of studies have been published to prove or disprove this or that theory. My goal is to point out previous valid studies that I have first-hand experience with that may have been overlooked today because "most people" have been conditioned to believe it is not important.
There is no way to convince anyone about psychoacoustic phenomenon because every person experiences it differently. All one can do is explain the physics and realize that there is a high probability some facet might be audible. Just because "most people" agree on a particular topic in such a subjective area does not mean all other observations should be abandoned. Stop picking on those of us in the minority.
I use crossovers without group-delay distortion simply because of three reasons:
1.) I know how to do it
2.) There is still no definite answer about how and when group delay distortion is audible, so why not try to be on the safe side ?
3.) Technically speaking it is a well-known fact that transmission channels with group-delay distortion have a lower information capacity than those without group-delay distortion.
Therefore I take care of this.
Regards
Charles
1.) I know how to do it
2.) There is still no definite answer about how and when group delay distortion is audible, so why not try to be on the safe side ?
3.) Technically speaking it is a well-known fact that transmission channels with group-delay distortion have a lower information capacity than those without group-delay distortion.
Therefore I take care of this.
Regards
Charles
Last edited:
I've had a browse through that article, and with all due respect to the author, it seems like he's only just discovering the consequences of crossovers, phasing, lobing etc for the very first time in that project, and basing everything on what sim software spits out. (I didn't see anything in the article to suggest that the proposed design had been built or measured)
In particular the following bits made me doubt the authors experience with the matters we're discussing.
Referring to the effects of different path distances to each driver off the vertical axis:
Basic stuff.
Despite simulating and graphing everything he doesn't seem to understand the relative phase of the crossover outputs of a 4th order L/R, and why they should track with frequency.
Referring to the interpretation of on and off axis frequency plots vs polar plots:
Again the author seems genuinely surprised at how the two ways of looking at the same information (FR plots at different angles, vs polar plots at individual frequencies) relate to each other, and it's pretty clear that this is all brand new to him, he pretty much says so himself.
I really don't want to appear to be hammering on this guy, because I'm not. Everyone has to start somewhere and I commend him for going through the motions he has in trying to understand the various design tradeoffs and what's going on. But it's clear that it's a simulation exercise and his practical and theoretical grounding on what's going on behind the veneer of the simulation software is quite shaky, so I really don't think this article could be considered a suitable reference to prove the issue one way or the other.
In his graph showing the vertical lobing for various filter orders, he shows the lobe downwards for both 1st order and 3rd order butterworth, in direct contradiction to one of your previous references where the 1st order lobe is shown tilting down, and 3rd order tilting up:
Power
Why could this be ? Lets look at the accompanying text:
It seems likely to me that for the 18dB/oct filter he has simulated it with the tweeter in reverse phase - probably based on advice in LSPCad, and the naive assumption that the phase shift must be +90 degrees. I suspect he's unaware of the fact that a -90 degree phase shift (achieved by leaving the tweeter in phase) will also sum flat on axis but tilt the lobe in the opposite direction.
A quick question, have you built an 18dB/oct crossover before and measured the effects on lobing, or are you just basing your position on simulations ? (That's an honest question, not meant as a jab)
Nope, nothing like that going on.
Last edited:
It's only the correct choice if the signal remains in the electronic domain.
As soon as the signal enters the acoustic domain, via two or more non-coincident, non-ideal point source drivers, it's no longer clear cut.
Now you have to weigh up a number of different trade-off's, where a more theoretically "perfect" filter often makes other parameters that are arguably more important get worse, sometimes much worse.
For example a summed 1st order filter is perfect in every way in the electrical domain, and is the only type to be able to reproduce a square wave.
However as a crossover for a speaker it is severely limited in all but a few cases, due to beaming from huge overlap, high distortion from drivers being driven beyond their useful working range, and incomplete summing due to very few drivers having enough overlap to sum flat over such a wide range with a 1st order filter.
Another example is that a 4th order L/R filter has many desirable properties on paper, but one of it's limitations is that the power response is 3dB down at the crossover frequency - something that only manifests itself when the summing is done in the acoustic domain using non-coincident drivers.
So I would say that as far as getting a real world speaker sounding the best "The configuration with the least amount of group delay is the correct choice, regardless of driver relative polarity." does not necessarily hold.
More often than not, filters with "worse" characteristics are chosen to optimize other parameters of the speaker in the accoustic domain. For example using steeper filters to better roll off drivers and reduce overlap, despite the "worse" phase and impulse characteristics of such filters.
Even if you reach the point where the non-ideal characteristics of the filters such as impulse response may become audible, it could still be the lesser of two evils when the alternative could be high distortion from a driver etc.
Assigning relative importance to different trade-off's and flaws is what speaker design is all about, and what leads to the vast array of different designs that are out there.
I presume the two graphs with driver polarity alternating vs not alternating were to make the alternating case look worse for phase shift and group delay, however I can't help noticing the substantial ripple in the amplitude response in both cases.
You don't state the filter configuration but I'm going to guess that it's a parallel topology ?
If so you may or may not be aware that an 18dB/oct 4 way crossover can't be satisfactorily implemented in a parallel topology without severe errors in the response - the crossover frequencies are too close together.
Linkwitz has a good summary on this near the bottom in the section crossover topology mistakes:
Woofer crossover & offset
So we can't really tell much in this example about subtle variations in group or phase delay through the crossover regions, when both filters are flawed in their implementation.
DBMandrake :
The implementation is indeed a cascade configuration, HF in first , the LP filters feeding the next lower section. Sorry for the confusion about amplitude ripples. Iin the NORM simulation, the xover freqs are slightly staggered to minimize residual phase errors from the adjacent sections. Note the dip in total power is only ~ -0.5 db. One argument that odd order BUT is better than 2nd/4th LR is based on the fact that LR has larger total power dip at xover, and that the power dip is indeed audible. (This may be a matter of opinion, I believe it to be true).
Forgot to reset xover freqs for the "all driver in phase" run, so there are ripples in phase and GD.
HOWEVER, the total amount of phase shift and GD is still correct for each respective case. With all drivers same polarity, there is twice as much phase shift / GD. Same is true for 4th LR with same polarity drivers.
Here are new sims with xover freqs same:
A you can see, the amplitude response is even worse for both cases if the xover freqs are not staggered a little.
This interaction is not peculiar to BUT, but occurs in all xover topologies, the residual phase errors increasing as filter order increases. With 4th LR, the interaction will be even more severe.
Measurements on the actual implementation of this configuration yield identical results for amplitude and phase, so I know for certain that this sim is accurate and that the implementation is indeed a cascade configuration.
FYI, I had already discovered that cascading the sections was optimum several years before Linkwitz/Riley published the references you mentioned. I did my work in the late 70's. I was pleased to find someone agreed with my findings.
The implementation is indeed a cascade configuration, HF in first , the LP filters feeding the next lower section. Sorry for the confusion about amplitude ripples. Iin the NORM simulation, the xover freqs are slightly staggered to minimize residual phase errors from the adjacent sections. Note the dip in total power is only ~ -0.5 db. One argument that odd order BUT is better than 2nd/4th LR is based on the fact that LR has larger total power dip at xover, and that the power dip is indeed audible. (This may be a matter of opinion, I believe it to be true).
Forgot to reset xover freqs for the "all driver in phase" run, so there are ripples in phase and GD.
HOWEVER, the total amount of phase shift and GD is still correct for each respective case. With all drivers same polarity, there is twice as much phase shift / GD. Same is true for 4th LR with same polarity drivers.
Here are new sims with xover freqs same:
An externally hosted image should be here but it was not working when we last tested it.
An externally hosted image should be here but it was not working when we last tested it.
A you can see, the amplitude response is even worse for both cases if the xover freqs are not staggered a little.
This interaction is not peculiar to BUT, but occurs in all xover topologies, the residual phase errors increasing as filter order increases. With 4th LR, the interaction will be even more severe.
Measurements on the actual implementation of this configuration yield identical results for amplitude and phase, so I know for certain that this sim is accurate and that the implementation is indeed a cascade configuration.
FYI, I had already discovered that cascading the sections was optimum several years before Linkwitz/Riley published the references you mentioned. I did my work in the late 70's. I was pleased to find someone agreed with my findings.
Last edited:
Hmm, I'm not sure that you mean the same thing as Linkwitz when you say "cascade configuration". His description I linked to refers to tapping the input for the next bandpass section up in frequency from the output of the high pass section of the previous one.
Is your software simulating this ? There shouldn't be residual phase errors and amplitude ripples with that topology, (that's the whole point of it) at least in an active design.
I've also seen conventional parallel connected bandpass filters referred to as "cascade" filters if they are constructed from a high pass followed by a low pass, rather than as a transformation bandpass. I wonder if this is a source of confusion between us.
I completely agree on that point, subjective as it might be. I think the power dip is very audible in any room with reflections. (At least for midrange/treble crossovers, maybe not as noticeable for bass/midrange crossover points)
I've tried listening tests years ago with the same drivers (a full range driver and a ribbon tweeter) with both 3rd order butterworth and 4th order L/R at 4Khz, comparing them both to each other, and to the unfiltered full range driver on it's own (as a reference) and I wasn't happy with the results of the L/R.
Even though the on axis result measured the same between the two different filters the subjective impression in the listening room was that the L/R sounded very sterile and "dead" without good focus and imaging, compared to either the full range driver on it's own or the 3rd order filter.
What sweet spot there was was very narrow in the vertical plane, and quite hard to discern.
By comparison the 3rd order filter sounded very close to the full range driver on the own, (at least through the crossover region, not in the high treble where the full range driver was falling off) and I think it's reasonable to suggest that's due to the flat power response. (Or more precisely the filter isn't adding additional loss to the power response over the drivers own natural characteristics)
A few days ago while searching for a reference for another discussion thread I came across an interesting article on the subjective sound of different crossover types where they crossed over two identical vertically mounted small full range drivers with various crossover orders, comparing them to one of the drivers on it's own unfiltered as a reference:
http://www.ctc-dr-weber.de/speaker/Audibility of different crossover types.pdf
I was quite interested to see that although subjective, what they found agreed completely with what I'd found years ago - of all the filter types they tested the on axis sound in a normal listening room the 3rd order butterworth sounded the closest to the full range driver by itself with crossover frequencies >=2Khz, with a slight edge to the in phase connection. They also found that at much lower frequencies (bass/midrange crossovers) that it didn't matter nearly as much.
The main problem with a 3rd order filter is the 3dB off axis "peak", but that only occurs if both drivers have extremely wide dispersion at the crossover frequency. With a real midrange driver, especially a largish one, the off axis fall off at the crossover frequency can be contrived with the right driver spacing and crossover frequency to counteract and avoid the off axis peak.
On axis the 90 degree phase shift keeps the response from peaking, vertically off axis in the "peaking" direction the fall off of the response of the larger driver reduces the response below the theoretical +3dB, fairly close to 0dB if you get it right.
Whether the drivers are connected in phase or out of phase matters here because the rolling off off-axis response of the larger driver will add extra phase lag compared to an ideal point source, and that will either help or hinder the phase shift caused by the distance differential...
(The driver phasing affects the horizontal off axis summing as well, as the increasing off axis phase lag of the midrange driver adds or subtracts from the relative phase)The end result if done well is a wide but fairly uniform vertical lobe which doesn't peak, despite a relatively large driver and large driver spacing.
So a 3rd order crossover can be made to work very well with large drivers with moderately wide spacing (~1 - 3 wavelengths) but for small closely spaced drivers with very little off axis fall off you'll get peaking problems so the 4th order L/R may be better there...
Has it occured to anyone that the narrow vertical sweet spot is due to fact that no matter which way listener moves, response is more rapidly approaching an infinite dip? With 3rd BUT, moving other way response slowly approaches only +3 dB, so sweet spot appears to be larger vertically.