My design was LR4, fc=1550Hz. I'm not deaf to it but I found the difference much smaller than anything else speakers do, and hence the last thing to optimize. I would certainly never try to improve overall phase at the expense of any other performance parameter. It's worthwhile only when everything else remains the same.
The difference was most audible on things like fast guitar riffs where the notes separated better temporally. I'm told the differences at low frequencies are greater, but I haven't been in a position to try that yet.
The DLCP does have electronic delay of course. So you can always phase match the drivers correctly. The story about IIR filters correcting phase does not include time of flight, you do need to correct that separately.
The difference was most audible on things like fast guitar riffs where the notes separated better temporally. I'm told the differences at low frequencies are greater, but I haven't been in a position to try that yet.
The DLCP does have electronic delay of course. So you can always phase match the drivers correctly. The story about IIR filters correcting phase does not include time of flight, you do need to correct that separately.
I basically agree, though my stance is a bit more moderate. Optimizing driver placement and delays for the front wave shifts the time misalignment to the rear wave and doubles it. That's fine for closed and vented box speakers. It isn't necessarily such a great idea for dipoles as the rear wave's generally not attenuated much by the room. So open baffle makes the design tradeoffs between excursion, group delay, and delay equalization a bit stronger. It can be attractive to accept the greater group delay of steeper crossovers (and steeper roll off of dipole equalization) in exchange for more consistent directivity and more linear driver operation. Particularly if the application doesn't require video sync so that the increased latency of using some form of inverse allpass to hammer the group delay back to flat isn't much of a concern.
ASIO learning curve aside, Phase Arbitrator makes it trivial to A/B min phase against linear phase. My experience with several different speakers I've worked with is the audibility is roughly constant across the spectrum. I do have some difficulty hearing phase corrections in the deep bass---<40Hz---but my hearing's not particularly good there anyway.
ASIO learning curve aside, Phase Arbitrator makes it trivial to A/B min phase against linear phase. My experience with several different speakers I've worked with is the audibility is roughly constant across the spectrum. I do have some difficulty hearing phase corrections in the deep bass---<40Hz---but my hearing's not particularly good there anyway.
If you adjust timing by physically shifting the drivers you get that, yes. I'd propose aligning the drivers such that the phase shift in front or in the back is the same and correct the remainder electronically. That way the electronic time alignment is valid for both front and back waves.
The extra time delay that one dials in is actually a good indication of whether the driver responses have been hammered into the right filter slopes. If this is so (that is, to well beyond the crossover frequency) the additional time delay required corresponds almost exactly to the horizontal position of one voice coil to the other.
Steeper filters with overall phase correction aren't free of faults. Since in an LR filter the phase responses of both drivers match, the phase correction filter actually turns both of them into linear phase HPF/LPF, with attendant preringing. When you go off axis (vertically) the pre- and post ringing of the filters no longer cancel, so a standing listener gets pre-ringing at the crossover frequency. It may not yet be a serious sonic issue with 6th order filters but it's visible in impulse response plots.
The extra time delay that one dials in is actually a good indication of whether the driver responses have been hammered into the right filter slopes. If this is so (that is, to well beyond the crossover frequency) the additional time delay required corresponds almost exactly to the horizontal position of one voice coil to the other.
Steeper filters with overall phase correction aren't free of faults. Since in an LR filter the phase responses of both drivers match, the phase correction filter actually turns both of them into linear phase HPF/LPF, with attendant preringing. When you go off axis (vertically) the pre- and post ringing of the filters no longer cancel, so a standing listener gets pre-ringing at the crossover frequency. It may not yet be a serious sonic issue with 6th order filters but it's visible in impulse response plots.
I've been doing what you suggest for as long I've been building dipole speakers. 
In practice most of the preringing I measure is from DAC filters---CS4272 is the main part I've worked with, which outputs quite symmetric sync functions when fed an impulse---and not preringing from time reversed IIR. DAC ringing dominates up to maybe60 degrees off axis even with multiple time reversed IIR passes correcting different effects. Beyond 60 degrees SPL drops off rapidly due to the dipole null so it works out OK. Wouldn't surprise me if reduced preringing sounds better but I don't yet have a DAC where I have control over the on chip filtering.
The data I have is anecdotal but suggests significant performance differences between delay equalization solutions. FIR appears to be worst, FFT hit and miss depending on how good the implementation is about windowing, and time reversed IIR most robust---though I'm only aware of two commercially available time reversed IIR implementations. Inverse allpass IIR delay line might outperform time reversed IIR as it's operating with a shorter "block size" but I suspect that may get swamped by residual group delay ripple. Be interesting to see how the performance turns out.
In practice most of the preringing I measure is from DAC filters---CS4272 is the main part I've worked with, which outputs quite symmetric sync functions when fed an impulse---and not preringing from time reversed IIR. DAC ringing dominates up to maybe60 degrees off axis even with multiple time reversed IIR passes correcting different effects. Beyond 60 degrees SPL drops off rapidly due to the dipole null so it works out OK. Wouldn't surprise me if reduced preringing sounds better but I don't yet have a DAC where I have control over the on chip filtering.
The data I have is anecdotal but suggests significant performance differences between delay equalization solutions. FIR appears to be worst, FFT hit and miss depending on how good the implementation is about windowing, and time reversed IIR most robust---though I'm only aware of two commercially available time reversed IIR implementations. Inverse allpass IIR delay line might outperform time reversed IIR as it's operating with a shorter "block size" but I suspect that may get swamped by residual group delay ripple. Be interesting to see how the performance turns out.
By off axis I was referring to vertically i.e. where the time alignment no longer works. It depends how steep the filters are whether or not pre-ringing is objectionable. But unless the drivers already have a >2nd order roll-off near the crossover frequency, LR4 is a nice conservative choice.
Pre-ringing of DAC's is pretty much invariant of the chip type or brand. The filter is almost always a half-band type with 0.4535fs pass-band. Audibility of pre-ringing at >20kHz is debatable but I would be very wary of allowing it smack in the middle of the audio band.
I would be rather skeptical of rules of thumb saying one method of approximating a certain response is better than another, especially when those rules derive from experiments. After all, implementing the same filter using either FIR or time-reversed IIR does not change just one variable like a good experiment should. That said I think logic concurs with you.
Since a straight FIR takes most processing cyles (but the least memory) it is reasonable to expect implementations to skimp on the length of the filter kernel. For my 1550Hz / LR4 filter, a short FIR filter is accurate down to 28 bits (coefficient length...), so an alternative implementation would offer no benefits. At 155Hz things would look decidedly different.
Time reverse is very cheap in terms of processing cycles so there what matters is having enough memory. On a PC you can implement a time-reverse filter virtually flawlessly with hardly any effort. Only when latency is an issue should any PC implementation be of a different type.
I would classify tapped delay + IIR (basically the only option left when you have a stream-based DSP) as the most troublesome in terms of optimization.
Pre-ringing of DAC's is pretty much invariant of the chip type or brand. The filter is almost always a half-band type with 0.4535fs pass-band. Audibility of pre-ringing at >20kHz is debatable but I would be very wary of allowing it smack in the middle of the audio band.
I would be rather skeptical of rules of thumb saying one method of approximating a certain response is better than another, especially when those rules derive from experiments. After all, implementing the same filter using either FIR or time-reversed IIR does not change just one variable like a good experiment should. That said I think logic concurs with you.
Since a straight FIR takes most processing cyles (but the least memory) it is reasonable to expect implementations to skimp on the length of the filter kernel. For my 1550Hz / LR4 filter, a short FIR filter is accurate down to 28 bits (coefficient length...), so an alternative implementation would offer no benefits. At 155Hz things would look decidedly different.
Time reverse is very cheap in terms of processing cycles so there what matters is having enough memory. On a PC you can implement a time-reverse filter virtually flawlessly with hardly any effort. Only when latency is an issue should any PC implementation be of a different type.
I would classify tapped delay + IIR (basically the only option left when you have a stream-based DSP) as the most troublesome in terms of optimization.
Last edited:
Agree with your observations on implementation tradeoffs and classification of tapped delay + IIR. Thanks also for the data point on your 1550Hz LR4; I've been meaning to give something similar a whirl but haven't gotten to it yet.
My experience is there's enough other time alignment stuff going in the vertical plane DAC and DSP preringing isn't the dominant effect. Volume displacement requirements for dipoles are large enough the vertical time of flight across the driver arrays of even fairly compact designs is a couple milliseconds. That swamps the DAC or time reversed IIR preringing I've measured. Sacrificing dipole bass response or switching to closed box does, however, reduce the time of flight to roughly the ~600ms preringing duration. But the temporal bleed from preringing is still something of a wash with respect to the time spread from, say, first reflection off the ceiling to the listening position. It's been my experience from switching drivers and time reversed IIR on and off that the time of flight spread in first reflections can be more subjectively significant than the preringing in direct or reflected sound. The reflections are, after all, often within a few dB of full SPL rather than 30 or 40dB down as preringing tends to be. I'm not saying preringing in DSP and DACs isn't worthy of attention. Just that, within the overall design, optimizing driver placement or installing a ceiling absorber can be more useful.
Yeah, DAC filters are pretty much the same---I've been thinking of bypassing them (though parts like the ESS DACs which accept arbitrary coefficients are an interesting alternative). Not sure if it will go anywhere but it seems like the sort of thing where predicting results a priori is tricky enough it's worth giving it a try. Working with LR4 or steeper crossovers makes bypassing fairly straightforward for subs, woofers, and most mids. Avoiding aliasing is more complicated for tweeters but not unreasonable for scenarios such as upsampled redbook audio. Sample rate conversion to reduce filtering costs by aligning Fs with drivers' operating frequencies is attractive too, though not an option with the TAS3108.
My experience is there's enough other time alignment stuff going in the vertical plane DAC and DSP preringing isn't the dominant effect. Volume displacement requirements for dipoles are large enough the vertical time of flight across the driver arrays of even fairly compact designs is a couple milliseconds. That swamps the DAC or time reversed IIR preringing I've measured. Sacrificing dipole bass response or switching to closed box does, however, reduce the time of flight to roughly the ~600ms preringing duration. But the temporal bleed from preringing is still something of a wash with respect to the time spread from, say, first reflection off the ceiling to the listening position. It's been my experience from switching drivers and time reversed IIR on and off that the time of flight spread in first reflections can be more subjectively significant than the preringing in direct or reflected sound. The reflections are, after all, often within a few dB of full SPL rather than 30 or 40dB down as preringing tends to be. I'm not saying preringing in DSP and DACs isn't worthy of attention. Just that, within the overall design, optimizing driver placement or installing a ceiling absorber can be more useful.
Yeah, DAC filters are pretty much the same---I've been thinking of bypassing them (though parts like the ESS DACs which accept arbitrary coefficients are an interesting alternative). Not sure if it will go anywhere but it seems like the sort of thing where predicting results a priori is tricky enough it's worth giving it a try. Working with LR4 or steeper crossovers makes bypassing fairly straightforward for subs, woofers, and most mids. Avoiding aliasing is more complicated for tweeters but not unreasonable for scenarios such as upsampled redbook audio. Sample rate conversion to reduce filtering costs by aligning Fs with drivers' operating frequencies is attractive too, though not an option with the TAS3108.
Not yet decided, we still have a couple of sets available. If you are interested please drop Jan-Willem (at) hypex (dot) nl an email.
First pictures running system from a beta tester:
forum.zelfbouwaudio.nl • Toon onderwerp - Hypex 12 kanaals gebalanceerde pré met dsp
In dutch, but you can see the dsp board, input board and more.
Regards,
Gerhard
forum.zelfbouwaudio.nl • Toon onderwerp - Hypex 12 kanaals gebalanceerde pré met dsp
In dutch, but you can see the dsp board, input board and more.
Regards,
Gerhard
My conclusion was flexible, real time linear phase procssing requires time reversed IIR and processing power on the order of an audio SHARC. A DLCP type platform equipped with this would be compelling as the end user cost riser and hassle reduction from moving from the TAS3108 to something like an ADSP-21366 is lower than implementing the equivalent on a PC and much lower than the VisualDSP++ licensing cost and time needed to program the SHARC. A lower cost, complexity, and playback latency alternative is to move delay equalization offline by performing it as part of ripping. Linear phase support for streaming media is sacrificed but that's an acceptable compromise for many stereo playback applications. A DLCP or similar can then be used for forward time processing.