Elliptical Filters Revisited

[villastrangiato]

I've been looking into crossover options for a design involving the new BG Neo8-S - Jon's elliptical filters look very intriguing but the group delay possibilities are a little scary - anyone here have a chance to get critical listening time in with either his Modula or M8ta designs to get detailed impressions of how they handle voices in the critical range from 1k to 2k? I'm curious if anyone has done a comparison in these designs with more conventional L-R or Butterworth versions.

Also, if Jon cares to, I'd love to hear his opinions about proper criteria for setting the resonance of the 'tank' circuit portion of the filter - specifically what some of the limitations might be in terms of how close one can get to the crossover region - and the resultant impact on ripple and phase linearity.

[Amphiprion]

I do not think the group delay/phase shift should be of concern to you.

If you would like to try an experiment, here is one that I posted some time ago.

http://mark.hayenga.com/audio/phase_shift/index.htm

[villastrangiato]

Thanks for the info Mark. I'm not too sure whether I should be concerned or not. I can't say I've heard the effects of group delay. I've seen a fair amount of discussion over the years about it - with a large number of people claiming they can hear ringing and transient differences between a fourth order L-R and first order network. I recently picked up a BBE DS48 to try a number of combinations back to back to see if any differences could be gleaned. My guess is if the slope is steep enough, it's debatable how audible a high rate of phase change will be over a 1/4 octave. Given that I've not yet heard an elliptical filtered crossover yet - it's guesswork for me. Additionally, for many, with a poorly implemented first order system, the resulting beaming of a large format driver operating at frequencies it shouldn't would probably be more noticeable than the effects of group delay in a steeply filtered variant.

The technical papers on elliptical filter papers I've seen thus far suggest that the high levels of group delay are indeed audible and a point of concern - but I haven't yet confirmed this with anyone who owns such a system. Additionally, there may be a difference in the way Jon implements the design that mitigates the high group delay side affect.

I'm not sure we're talking about the exact same thing though. The issues are the rate of phase change with respect to frequency and degradations of transient response and damping - not simply the "phase delays" you seem to be referring to.

[Amphiprion]

The examples I posted are exactly that - increasing group delay (around the crossover point) as well as increasing total phase shift magnitude. They are just Linkwitz-Riley analogues, not elliptical. Look at the frequency magnitude plots. Frequency is phase is group delay (in an LTI system, anyway). I sort of specialized in DSP when I did my bachelors in electrical engineering, so I'm well versed in this.

I could whip up some all-pass filter simulations for an elliptical filter if you would like. Specifically to test for group delay alone. Anything you want, really. Which of those tunes was your favorite, and what type of filters would you like made? I'll generate all the files, plots, and code for anyone to double-check me and make another little blind test just for you I need to brush up on my Matlab skills anyway.

[twest820]

If it helps I personally find the difference between a phase corrected LR4 and a regular LR4 audible and prefer the corrected version. I haven't listened to eigth order crossovers but it's not hard to grab Thuneau's Phase Arbitrator or Frequency Allocator and A/B the phase correction to see what you like. My experience with LR4 is it's definitely nice but not a major issue. In addition to the narrower transition band an eigth order is in phase at the crossover frequency so I'd expect it to be more liveable than LR4 alignments which go in phase - antiphase - in phase rather than in - anit - in - anti - in.

Got a link for the Neo8-S? I don't see it at Parts Express and BG doesn't have anything on their site.

[villastrangiato]

Sounds intriguing Mark. But before you go through the trouble, it might be instructive to know how the filters you've tested are being implemented - specifically if they are active or passive. It's one thing to simulate a transfer function with op amps, caps, and resistors - or with digital conversion and processing - quite another to approximate them with passive components and the inevitable complications relating to interactions with drivers and loading from one stage to the next. I'm a big fan of active crossovers with all the inherent advantages (insertion loss, saturation/compression....etc...), but that's not what I'm working with.

[villastrangiato]

Got a link for the Neo8-S? I don't see it at Parts Express and BG doesn't have anything on their site.

I just found out about it a few weeks back talking to Igor at BG. He says it's available directly from BG as is the NEO10. According to him, it's essentially the same as the current NEO 8 but with thicker magnets and a slightly wider gap - providing higher power handling and output - should be interesting. 8)

[villastrangiato]

If it helps I personally find the difference between a phase corrected LR4 and a regular LR4 audible and prefer the corrected version. .

A "phase corrected" 4L-R? Could you explain that a bit? I haven't run into anyone yet who "corrects the phase" of a Linkwitz-Riley crossover. If you overlap corner frequencies or apply an allpass delay - you're not really doing anything to reduce the actual rate of phase change with frequency - just possibly reducing a little bit the band of frequencies over which the signal is "smeared" in time and paying for it with arguably noticeable magnitude anomalies. With the higher order (more underdamped) crossovers, you might be able to reduce the audibility of ringing and tonal shift at the edge of the crossover band with very steep slopes - but somewhere in the middle - you're still dealing with the potential of heightened phase distortion smack dab in the middle of a critical range of hearing.

[cjd]

The only way I've ever heard perfectly phase corrected crossover was active - it was done in the digital realm. It was not a Behringer, I don't recall what brand it was though.

I FAR preferred the passive Dennis Murphy (I think - this was a Salk speaker) crossover on the same speakers.

Hash and distortion from cone break-up is, to my ears, far worse for the sound than a steeper crossover filter. You can not isolate JUST the difference between a 1st order and an 8th+ order elliptic filter because you can't work with drivers that are perfect with even distortion over the entire range. So many comparisons include softer drivers that inherently damp more cone break-up so you can cross lower order and not have it be painfully obvious - and a lot of people really like that soft cone mush (just like many younger folks now prefer the sound of low bitrate MP3 hash )

[villastrangiato]

..... and a lot of people really like that soft cone mush (just like many younger folks now prefer the sound of low bitrate MP3 hash )

:lol: Very true...

I've been wondering for quite some time now how long it's going to take for a good FIR crossover that doesn't cost 6K to hit the market.

http://www.deqx.com/products.php

But then, I've also been wondering for a little while now (about 20 years.... :roll: ) how long it's going to take for a mainstream electronics manufacturer to produce amps and preamps for home use that use balanced inputs/outputs.

[cjd]

Oh yeah. It was the DEQX I heard. So I don't know if it was the implementation of the filters themselves, or something inherent in how they were done, that sucked the life out of the music for me.

[villastrangiato]

You're not the first person whose had that impression. I haven't had a chance to hear it yet, but on the surface, I'd have to say - why would anyone consider a 300db per octave slope necessary or even desirable - wouldn't it make more sense to apply the majority of sampling/processing/storage/reconversion resources to achieving a more modest result - say a 100db per octave slope - you'd still be in "beyond audibility" territory with respect to stop band and phase artifacts. The issues concerning impulse response "pre ringing" that seem to plague current approaches remind me of the myopic approach taken back in the early eighties at the dawn of consumer digital audio (wow, I suddenly feel very old... 8O ) where brick wall anti-aliasing filters were commonplace before someone (Philips) finally realized dedicating oversampling resources to the brick wall filtering problem was a much better approach. It's funny, just because you have high frequency digital circuits at play, designers seem to think the laws of physics don't apply anymore. Extremely sharp changes in energy - whether electrical or mechanical are almost always accompanied by instability and uncertainty. Anyone familiar with the Heisenberg uncertainty principle would probably see strong parallels in the comparison of IIR and FIR approaches. Perfection shouldn't be the goal - just the limits of audibility. It's been done with compact discs - why not digital crossovers? The double sampling rate suggests to me that the current SHARC 32 bit processor implementation is not up to the real demands that are being imposed by absurd 300 db/octave roll offs - just my hunch. For such an expensive system, I would expect at least 4 times oversampling to go along with the 32 bit precision.

[Dennis H]

The total phase shift of a Cauer is the same as the underlying topology without the extra notch filter. In Jon's examples, that's 4th order or 360 degrees. Adding the notch adds some phase wiggles in the stop band but hopefully the magnitude is down far enough that those aren't audible. The pic shows an LR4-1000 with steep notches an octave above or below.

[villastrangiato]

Thanks for the graphs. Some simple math already revealed the quality factor of 20 - I was just a little curious if this Q was derived from a software program or tables or if the designer chose it specifically for some reason. Jon hinted at a comparison to LR in the original MTM thread - so I was expecting a much lower Q. As for the absolute phase - I don't think anyone would argue that the phase shift itself is audible or even the highly attenuated magnitude response ripples just outside the crossover region. The essential bone of contention seems to be - how audible is the abrupt rate of phase change within the range of 1khz about the crossover point. A derivative with respect to frequency of the phase plots shown - in comparison to the same derivative for the crossover without the additional tank capacitor (lowpass side) would be much more instructive as to the actual group delay differences - should they exist. Regardless, I have to believe that Jon's Modula crossover design couldn't possibly have any serious audible detrimental effects in the crossover region - otherwise someone (namely Jon) would have acknowledged this already. Perhaps there are some effects which are minimal - and these are overstated by the general technical papers on Elliptical filters I've read thus far. And perhaps the choice of Q or the positioning of the tank circuit's resonance outside the crossover region has an impact on mitigating these issues - that's essentially what I'm trying to determine here.

[twest820]

A "phase corrected" 4L-R? Could you explain that a bit?

Well, corrected maybe isn't the best diction. The usual approach is to synthesize a time reversed IIR filter which converts another filter to linear phase---no shortage of papers on the subject if you do a search. If that's not how Jan implemented Phase Arbitrator he did something equivalent.

I just found out about it a few weeks back talking to Igor at BG. He says it's available directly from BG as is the NEO10.

Interesting indeed. :B I tossed BG an email to find out quantities, pricing, and specs and got a response saying the drivers should be released for sale in Q1 CY2010. No pricing or specs yet.

[villastrangiato]

Basically, for the purposes of a typical steep passive filter implementation, I don't see the use of all pass phase delays being very practical. My hunch is they aren't very common for a variety of reasons, not the least of which being cost and the limited benefit they would provide. In going back over the Modula MTM threads, the only group delay information I found was a "modeled" response. Perhaps to settle the issue, high performance crossover designs such as the Modula should be accompanied by Cumulative Spectral Delay plots and step response plots for proof positive that group delay design goals have been met. It's not as if only testers like John Atkinson provide such data in their evaluations. John Krutke, an individual more closely linked with this forum also makes extensive use of CSD plots and graciously provides them to forum members and others in the public domain. While I recognize this isn't exactly Jon's "day" job - some acknowledgement should be given to the fact that some of his designs are very popular and a more complete rendering of test data would be helpful in strengthening the broad acceptance and appeal they have achieved. As a side benefit, I believe it would provide more assurance for those toying with the idea of trying one of his designs to take the plunge and lay down the significant chunk of additional change to complete a more rigorously engineered project.

[PhilDSP]

As for the absolute phase - I don't think anyone would argue that the phase shift itself is audible or even the highly attenuated magnitude response ripples just outside the crossover region.

There may be a point being missed with such tests. In real listening situations with a passive filter, the impedance changes due to changing voice coil temperatures and different power demands for each possible frequency are going to make the phase shifts and group delay anomalies a moving target which will most certainly be audible if not necessarily recognizable IMO.

Neither are active filters a particularly good solution IMO because you just swap some problems with passive components for potentially greater problems and make it all the more harder to integrate well with the drivers and cabinet. Unless you a have a super sophisticated and powerful DSP system with the tools to engineer it properly of course.