Beyond the Ariel

Henckel said:
Since no speaker has a flat response down to DC ( 0 Hz) the impulse response has to have as much(Area) below zero as above zero.

Correct.

There's a good point about the energy allocation of an impulse - the entire spectrum of the signal is tilted towards HF, thus what you see is mostly HF artifacts. The reverse is true of a step function, which has a spectrum tilted towards LF, so you mostly see the highpass function of the woofer (which is difficult to measure except in nearfield).

Also, like an oscilloscope trace, you can really only see about 30~40 dB of visual resolution, thanks to the linear display. That's one area where a CSD is more useful, since the display is logarithmic in the amplitude and frequency domains. I usually use a CSD and impulse display together, to get a sense of where the worst time-domain faults lie.

Still, I very much like the absence of smoothing in the impulse response - it seems like I always get a million reasons from speaker designers why I'm not intelligent enough to interpret anything but a "smoothed" display - which of course removes all of the bad parts of the measurement technique and small-scale faults in the loudspeaker. If there's a resonance at 2, 4, 8, or 16 kHz, well, I'd like to see it, not have it "smoothed" away.

Maybe I'm a little oversensitive, but I can usually hear problems with HF resonances, even ones I'm not supposed to be able to hear. That's probably why I didn't much care for the first CD players and their discs, which had far more severe HF ringing in the time domain than a good-quality tweeter of the day (back in 1982, when tweeters weren't very good). If a tweeter (typically a metal-dome) has problems around 22~30 kHz, even though I'm not supposed to be able to hear it, I usually don't like it - it'll sound "rough" and kind of sandpapery to me. Thus, I like to measure beyond 20 kHz, and use lowpass fliters that don't degrade the time domain with ringing artifacts.

The "rumbling" quality of some of these horns - Meyersound included - is rather striking. It's all too obvious the air-load of the horn is not damping the diaphragm, contrary to the claims of JBL over many decades.
 
The audibility of ultrasonic resonances (or high-Q digital-filter artifacts) deserves more discussion. Obviously, if there is no program content to excite the resonance in the first place, it can't be audible - or would it?

Reading Zaph's Web page, I thought that if the 3rd, 5th, or 7th harmonic lands on an ultrasonic resonance, there could be consequences lower down in the spectrum - maybe a very narrowband rise in distortion, only occasionally stimulated by certain types of program material, but lurking in the background all the time.

I suspect there are substantial differences in audibility of narrowband resonances depending on the auditioner. Horns and low-resolution digital systems (like Red Book) are well-known for high-Q narrowband artifacts - and as far as I can tell, quite a few people don't hear these artifacts at all, and others very strongly object and don't like them.

I couldn't accept first-general CD's at all, and didn't buy a player until the Philips 4X oversampling technology came out - even then, I always kind of thought of it as being a more durable alternative to prerecorded Dolby cassettes, but not a truly hifi medium. When the Apogee UV22 dithering and sample-rate-conversion technique became more widespread in the early Nineties, CD's started to sound more acceptable-sounding, although still not as good as my late-Seventies vintage phonograph. But records were becoming harder and harder to find, and I reluctantly made my peace with the medium, although never really liking the sound all that much.

When I first heard really high-quality 96/24 PCM demonstrated by Keith Johnson (who also made the recordings) at the CES and VSAC, I was surprised how little it sounded like 44.1/16 and how close it came to analog mastertapes (also demonstrated by Keith Johnson on Ampex studio recorders). The sound we had all accepted as "digital" was nothing more than the artifacts of a rather poorly implemented low-resolution PCM system - exactly the same thing that was said at the birth of Compact Disc in 1982. When you flip the switch for yourself, all doubt is removed - Red Book is a poorly designed system that never had "mastertape sound" in the first place - and still doesn't, 25 years later.

What I find even more surprising is some people can't hear the difference between 96/24 and 44.1/16 PCM, which I find little short of astonishing. To me, it's like the difference between Dolby-B Compact Cassette and a 15IPS mastertape - obvious and impossible to ignore. The low resolution is one thing - not desirable, but kinda-sorta tolerable - but the metallic-sounding artifacts in the 5~10 kHz region are much harder to accept.

This particular artifact is frequently the result of lazy do-it-by-the-applications-handbook analog-circuit design following the DAC converter chip - and here's just one example: analog-stage op-amps that are too slow by not one, but many orders of magnitude.

Matt Kamna and I measured slew rates well in excess of 1000V/uSec coming out of the Burr-Brown PCM63 converter, along with a flat RF comb spectrum going past 20 MHz - and visible to 50 MHz. But - the common op-amps in even very expensive (more than US$5000) players typically have no more than a 13V/uSec slew rate (Philips/Signetics 5534, 5532, and its successors).

Slewing in the lowpass filter results in the sample interval being the incorrect time duration, which then translates into HF distortion in the audio band. Sampling theory requires that the time duration of the individual samples remains precisely the same as the original sampling operation, otherwise nonlinear distortion will result when the signal is reconstructed. The severity of the digital-reconstruction distortion is a function of the loudness and HF content of the signal (a larger proportion of the signal is slewed, and improperly reconstructed by the lowpass filter).

Again, we come back to HF artifacts, which some people are not bothered by, and others object strongly. They bother me, and since I don't have to report to a boss or the AES, I can design for my own tastes, not a committee.
 
Two Measurements, Two Results

Here's my MLSSA Quad ESL57 measurement and the one posted upstream. This probably shows the value of the ultralow diffraction Spurlock Stand - much cleaner decay performance. I think this demonstrates the difference that measurement protocol can make - absence of ringing in the lowpass filtering (analog or digital, software or hardware), absence of peaking in the measurement microphone, and a low-diffraction stand. If you can't see it in the measurement system, what chance do you have of fixing it in the loudspeaker?

Then again, if the other measurement is a Quad ESL63, that's a sad comment on how much it went downhill. But I suspect measurement artifacts instead.
 

Attachments

  • 2-quads.jpg
    2-quads.jpg
    95.2 KB · Views: 1,131
I though I'd chime in on the impulse response since the impulse of one of my older designs popped up. The system had a low frequency cut off of 25 Hz and was flat out past 20 k Hz. Now if is true that any system that is not flat to DC must have equal area above and below the zero axis. BUT the positive area is that contained under the impulse (finite width). The area below the zero axis isn't always very obvious (as in my impulse) because if represents the decay of the energy of the impulse back to the equilibrium state. With a low, low frequency cut off this takes the form of a very long, and very low amplitude decay (talking in a simplified sense here). So while the area above and below the zero axis are the same, the area above is high amplitude, short time duration (area = A x DT) and the area below has amplitude much, much less than the positive impulse amplitude with a much, much longer duration. Thus, for systems with very low cut off frequency the negative area is not very apparent and can be of such low amplitude that it is below the noise floor. Thus in my impulse the negative swing is not clearly visible at all. What the impulse does tell you is that there is no crossover induced GD in the system. The system is minimum phase.
 
john k... said:
I though I'd chime in on the impulse response...

John,

Thank you for the lucid explanation. This is what I believed to be the case.

BTW, I am glad to see that the ICTA system is looking like it will see the light-of-day :)

Edward

P.S. I would have posted one of my transient-perfect impulse responses, but they were not available to me at the time - so I 'borrowed' the one that I remembered from your orginal "B&O concept" system.
 
diyAudio Chief Moderator
Joined 2002
Paid Member
Re: Two Measurements, Two Results

Lynn Olson said:
Then again, if the other measurement is a Quad ESL63, that's a sad comment on how much it went downhill. But I suspect measurement artifacts instead.

Don't know if it has hardware issues, its from Stereophile. Its ESL63 for sure. Ii just saw that it went visually negative also. ESL63 is widely regarded as time coherent and fast. I posted it to help stimulate commenting on the impulse questions.
 
mige0 said:
Hi



Very clean impulse response for three speakers working together in the plot of post #1898 .
Could we have a x-axes scale please?

Greetings
Michael


That plot, as I recall, was the first 3 msec of the impulse. Nothing much to see after that. Here is the reproduction of a 2k Hz square wave by that speaker.

An externally hosted image should be here but it was not working when we last tested it.


And at 500 Hz

An externally hosted image should be here but it was not working when we last tested it.
 
Henckel said:
Since no speaker has a flat response down to DC ( 0 Hz) the impulse response has to have as much(Area) below zero as above zero.

This is your answer.

No acoustic source can create a static change in the atmospheric pressure (not even the space shuttle when it enters the atmosphere - it creates a doublet), hence any and all acoustic signals (impulse response or whatever) must average to zero. If it doesn't then its a fake or it didn't come from an acoustic source unmodified by something electronic - or a pencil!
 
Hi Lynn, all

“If either Dr. Geddes or Tom have more recent impulse data, many of the readers (including me) would love to see it. Impulse data for horns and waveguides seems very hard to come by - you sure don't see it on the prosound sites, that's for sure - all those guys do are those 1/3 octave smoothed "happy graphs".”

Well, some measurements of a one I’m working on now are as “up to date” as I can come up with. It is a larger speaker than the SH-50 with 9 drivers but is along the same line as the SH-50 where all the drivers combine coherently and radiate as one source in time and space.

I would comment that the impulse response, while conceptually appealing, is probably not that informative so far as having some clue as to what you need to fix to make it better.
Actually, the magnitude and phase is better and also describes the impulse response.

Lets start with the speaker’s magnitude and acoustic phase, the speaker I’m working on is shown in the first response.
Unlike the SH-50 which has a passive crossover, It is an active (self amplified) speaker which has DSP based linear phase crossovers and has a 80Hz high pass included.
As it has been raining here, I have been doing the preliminary alignment and “learning the processor”, measuring in my living room so this is not the final setup, that has to be outdoors. These were taken about a foot and half from the mouth to reduce room contamination, in real life, this is done on a tower at two meters. One-tenth octave smoothing applied to magnitude & phase display, 20Hz bandwidth as I recall.

So, if one takes the magnitude and phase and does an Inverse FFT operation on that measurement, one gets the curve below which is the Energy vs Time Curve.
This shows the radiated energy in log mag scale (like the amplitude curve) vs Time, with the concept that all the energy is sent at one time but it is delayed reaching the microphone and spread out by the speaker.
Anyway, if you remove the Log scale and use linear, one has the Magnitude vs Time curve.

The energy is in an “envelope” vs time, that energy has a frequency dependant phase shift and so, can be described in 3 dimensions, Time on one axis, the Real or “in phase” energy AND the imaginary part which is the reactive plane 90 degrees away.
A better picture than my display and better explanation than mine of the Heyser Spiral is here around page 40.
http://www.aes.org/sections/pnw/reference/tef_man.pdf

Anyway, the Real or resistive phase portion of the energy in linear scale is the “Impulse Response” and the reactive part is the Doublet response. Both describe a valid view of the “elephant” but are 90 degrees apart.
The TEF does not store-smoothed data or allow for smoothing of data converted to another domain, here is the impulse response when the response data is smoothed to 1/12 octave.
I’ll post a measurement of the speaker when its finished if you’d like.
Hope that helps.

Tom Danley
 

Attachments

  • kydsl1.jpg
    kydsl1.jpg
    86.6 KB · Views: 990
Um, what is the benefit of smoothing an impulse response? I don't get it.

I thought the whole purpose of looking at an impulse response is to see artifacts that are in the time domain and see physical reflections, diffractions, and HF resonances "as they are", in the physical world. It's the direct equivalent of an oscilloscope - magnitude vs time, no processing. It doesn't correspond to hearing - true - but does correspond directly to loudspeaker defects.

The audibility of this or that artifact may be a matter of academic debate, but their existence isn't - that's real and shows up on the display. In the interest of the clearest, most artifact-free display, it does take a careful test protocol with a low-diffraction microphone that is free of ultrasonic peaks, a ADC conversion process that is likewise free of visible pre or post-ringing (implying a sample rate several times higher than the highest frequency of interest), and windowing profiles with low time artifacts (not a rectangular window, for example).

I'm a little puzzled by the conversion process of the TEF system. Is it incapable of making a true impulse response, and has to rely on a computation from the frequency and phase domain instead? Does this imply assumptions about minimum phase that may not occur in real loudspeakers? Not clear on the concept here, or what benefits TEF has over MLS measurements, at least for time-related measurements.

The impulse response we've seen of the ESL63 are frankly kind of bad - there are plenty of direct-radiators that are better than this, and original ESL57 is quite obviously better. I guess that explains the continuing popularity of the original, despite its many faults - low SPL capability, quite narrow lateral beaming, and somewhat tempermental room positioning requirements.

As we are seeing, loudspeaker systems that quiet down 35-40 dB in less than 500 microseconds are not that common. I have big reservations about electronic gizmos that purport to offer inverse compensation of time errors, since these same errors are typically are extremely fine-grained spatially - move the microphone (or listener) a few inches, and the time error can be radically different. Not necessarily any larger or longer, but a completely different fine-grained structure in the time domain.

This is a dead giveaway of diffraction or reflection, and a sign that electronic correction will makes things worse, not better, since the exact correction only works for a very small position in space, and makes things worse everwhere else. Errors like this cannot be corrected in either the time or frequency domain (well, you can, but it makes it worse), but have to be removed at the source.

To remove these errors at the source - the real, existing, physical world - you have to measure the precise distances and work out the likely location of the bad actor. Could be an internal cabinet reflection, could be a reflection off something on the front panel, could be a sharp edge somewhere, could be an unhappy driver resonance, you have to look, correct, and re-measure. Over and over again. If there are so many reflections (or resonances) all you see is a lot of random-looking clutter, well, then you know you have a basic design error - not only that, one that cannot be compensated for in the time or frequency domain.

This is kind of an old-school approach, and enters into discussions of whether or not XYZ fault is, or is not, audible, but I just don't like defects like this on principle. Now the arguments about audibility of phase distortion and crossover topology are a different thing entirely - that's the first part of the impulse, NOT the decay portion, which interferes with spatial and harmonic-decay impressions. In a feedback amplifier circuit, if you find even a miniscule tendency to ring at HF or approach oscillation under dynamic conditions, you don't debate audibility - you remove the defect and re-design as necessary. In the analog electronics world, that's just considered good design practice.
 
Hi Lynn

The TEF via TDS does not measure impulse response directly, it does measure magnitude and acoustic phase without regard for the item under test being minimum phase or not and does it well.
In the same way an impulse itself or a measure than generates one, can be used to get response and phase in the systems you are used to, mag and phase can be used to find the impulse response as well. I use the TDS because it is a very precise way to measure magnitude, phase and Time.

The TEF via MLS does measure impulse directly, except that process, like all the FFT and sequence based systems I have tried, it does not pass a “phase reality test” I devised, what they show is not actual acoustic phase or at least has a variable error.
Consequently, I would be skeptical that these systems give a “true impulse response” either while it does show both impulse and acoustic phase, neither are identical to a precise TDS measure of the same item.

Again, I am not sure how valuable the impulse response is in troubleshooting a speaker; it is only one view of a two dimensional event.
Take a moment and read the link, pages 39-41 about this, its pretty cool.
http://www.aes.org/sections/pnw/reference/tef_man.pdf
I mention smoothing the impulse as at 2KHz the raw data has 1/100 th octave resolution and 1/1000 octave resolution at 20KHz.

“I have big reservations about electronic gizmos that purport to offer inverse compensation of time errors, since these same errors are typically are extremely fine-grained spatially - move the microphone (or listener) a few inches, and the time error can be radically different. Not necessarily any larger or longer, but a completely different fine-grained structure in the time domain.”

Amen, that is a primary flaw of DSP applied to systems which have spatial problems in X and Y. Rarely if ever are the sources close enough together to add coherently like two woofers do side by side, they need to be less than a quarter wavelength more or less.
On the other hand, in hifi, spherical measurements which show lobes and “where the energy goes” are never shown (or done?) so such problems are unseen.
No one mentions the “perfect alignment” falls apart if you move up or down, right or left.

That kind of Steamroller DSP correction is not what your looking at here though, these are simple linear phase filters, no impulse correction via IIR or convolution etc although I know what you mean. My approach is to fix the source at the source with proper acoustic geometry.

This is what can be done with the ability to fine tune the time of arrival for two ranges by a fraction of a ms via time delay and having a linear phase crossover option.
The version using conventional Bessel filters wasn’t much worse fwiw.

There isn’t driver to driver interference anywhere because where each range overlaps, the acoustic dimensions are less than 1 / 4 wavelength across, the pressures sum together at a dimension that is acoustically “to small” to interfere and radiate fingers, beams or HOM’s.
In other words, where the mids take over from the compression driver, they are all less than a quarter wave apart (inside the horn) and away from the apex of the horn, they are a point source.
When the acoustic dimension is ”small enough”, multiple drivers cannot radiate as anything but “a point source in a conical horn throat” in there range, same for the woofers. All the drivers are tied together acoustically too, think about a horn, it is partly a “high pass” filter related to “how fast” its area expands. In a conical horn, it rate of expansion is very fast at the apex, it is a “high frequency” horn with poor lf loading as a result. If you tap in some distance away from the apex, one finds a slower expansion rate where mid drivers do feel horn loading, further down, one can tap in again with lower frequency horn drivers. It is assumed each set of drivers is appropriate for that range and circumstance.
A conical horn has in effect a variable “high pass” corner that depends on the physical position so the nub of the idea is to parse one single large horn into multiple operating ranges.
Each range is dispersed in depth, the hf driver is at the rear to the lf drivers at the front.
That depth is a time delay, which is incorporated in the crossover’s properties and sum into a “linear phase” like response.
The SH-50 you posted the impulse for is a three way speaker with 7 drivers and has a passive crossover.
(I was) Looking for some square waves, posted some SH-50 squares later in thread.
http://www.diyaudio.com/forums/showthread.php?threadid=71824&highlight=

I have found something that is of use at least casually at work evaluating our and others loudspeakers, someday (going into over the rainbow) I am told, we will have those wave files on the new web site.
Anyway, a generation loss demo might be of interest here in tracking down audible coloration. It tells you nothing measurement wise but magnifies what ever is wrong.
Take your measurement microphone and the speaker in question and hoist them on top of a tower about 15-20 feet (or more) off the ground with the mic about a meter away from the speaker. Play a recording of your choice and with a good recorder (24/96 or better pref), record the mic signal. Then, play the mic signal through the speaker and record a new mic track. Now, you have two generations of what ever is wrong, by three generations most speakers are pretty un-listenable with the subtle flaws now screaming at you.
If you have a speaker that can survive even four generations and still not be awful, you have a good speaker.
Do this in a standard room and location and now one adds the speakers interaction with the room (or not) into the audible result.
It is really interesting to hear a caricature of the speaker with all the warts progressively amplified via generation loss.
If your concerned about “digital” degeneration, use the second channel to do X stages of loss on the source too.
Anyway, I thought maybe that trick might be of use on your quest for a new speaker design.
Best,
Tom Danley
 
Thanks, Tom, most informative post. Compensating for linear delays is benign enough, and can be done passively with Nth-order Bessel lowpass filters. Be nice if Bessel highpass filters could somehow be persuaded to do the same thing, but that doesn't work.

Time errors, as this thread has shown, can really only be corrected in the time domain - frequency corrections can really mess things up even more - thus the necessity for making time measurements in the first place.

I surmise not all readers of this thread are aware that frequency measurements contain less information than impulse response - you can create frequency response, phase response, ETC decay, CSD waterfalls, and lots of other things from the impulse response, but the reverse is not true, since phase or time data is not present in a FR response plot.

One simple example are square and triangle waves, which have identical spectra in the frequency domain, with a diminishing series of 3rd, 5th, 7th and the rest of the odd harmonics. Although the magnitudes of the harmonics are precisely the same, the phases are different, making for a very different waveform.

Allpass functions are invisible in the frequency domain as well - we can argue about audibility of allpass functions in crossovers, or in drivers with peculiar time characteristics (the Lineaum tweeter exhibits allpass characteristics), but it's nice to measure first, then argue afterward. That's better than being in the dark and not knowing what's happening in the time domain - in effect, conceding the argument without bothering to take the data.

I think of that as the Ptolemaic theory of audio design - we already know the answers in advance, so we don't need to look into the stinkin' telescope!
 
I think the attached image can help understand the behavior of a system to a impulse. I used a Q =1, 2nd order HP filter in the example.

An externally hosted image should be here but it was not working when we last tested it.


First we must recognize that we are really dealing with a finite width pulse. In the figure I used a 0.1 msec pulse tp make things more obvious. The response of a (linear) system to a finite width pulse can be found analytically by obtaining the system response to a step input. The impulse response is then composed of the response to a positive step summed to the response to a negative step with the negative step delayed by the pulse width (figure should say red + green since green is already negative :)). In the figure the thin red line is the step response for the positive pulse. Since the system is a high pass filter the response consists of the initial rise to the step magnitude followed by decay. The decay starts immediately since the HP system can not sustain a DC level. In this example, the response to the positive step decays about 2 1/2 division in 0.1 msec. Then the negative sep response is added in and the result is the undershoot because the the magnitude of the step is the same but it starts from the reduce level due to the decay of the positive step.

Now, as the pulse gets narrower in time, the decay will not be as great before the negative step response is summed in so the undershoot won't be as negative. But as long as the system has a high pass nature there will always be some decay and some undershoot in the response.

If the system can support a DC level then there would be no undershoot and the sum of the negative step would return the level back to exactly zero.
 
diyAudio Chief Moderator
Joined 2002
Paid Member
Tom Danley said:
“I have big reservations about electronic gizmos that purport to offer inverse compensation of time errors, since these same errors are typically are extremely fine-grained spatially - move the microphone (or listener) a few inches, and the time error can be radically different. Not necessarily any larger or longer, but a completely different fine-grained structure in the time domain.”

Amen, that is a primary flaw of DSP applied to systems which have spatial problems in X and Y. Rarely if ever are the sources close enough together to add coherently like two woofers do side by side, they need to be less than a quarter wavelength more or less.
On the other hand, in hifi, spherical measurements which show lobes and “where the energy goes” are never shown (or done?) so such problems are unseen.
No one mentions the “perfect alignment” falls apart if you move up or down, right or left.

That kind of Steamroller DSP correction is not what your looking at here though, these are simple linear phase filters, no impulse correction via IIR or convolution etc although I know what you mean. My approach is to fix the source at the source with proper acoustic geometry.

It would be a blessing if geometrical & mechanical flaws causing time domain errors could be restored with post processing over a useful 3D projection space. That would be the pro sound speaker nirvana.
For absolute sound domestic audio, even if we had the DSP magic wand, we would soon pick up the strong signature of the DSP electronics with their trillion silicon devices micro world. An additional argument for killing the gremlins before they hatch.