ProSoundWeb Community

Please login or register.

Login with username, password and session length
Advanced search  

Pages: 1 [2] 3  All   Go Down

Author Topic: Trace Averaging Separation  (Read 825 times)

Michael Lawrence

  • Jr. Member
  • **
  • Offline Offline
  • Posts: 72
Re: Trace Averaging Separation
« Reply #10 on: March 07, 2019, 08:35:59 pm »

Interesting thread...with issues that often drive me crazy !

In my self-taught, measure-in-my-driveway/backyard type experience, I've come to a couple of simple preferences regarding averaging.

I try to average magnitude.
But only when there is high coherence, and response variations tend to be generally the same on and off axis.
So to keep it automated, I smooth measurements to taste before averaging, and the correct the average.
I do this driver by driver, assuming needed corrections are all minimum phase.
If corrections don't respond rather fully, I ditch them as non-minimum phase.
Hopefully I get smoother response optimized for a combination of on and off axis.. 

For phase ...no averaging...it seems to me phase is "inevitably to a spot"
So I use the previously corrected drivers' magnitude response, and phase align those responses to one point, at some distance that is close to equidistant to acoustic centers.

It's never measures as perfect as doing a mag and phase alignment to one spot,
but in today's world of easy to make damn near perfect response at a spot,
it seems like doing so may be its own new can of worms.  (at the expense of off axis)

I hope that made sense...please pass on any critique/observations/reservations on this methodology.


Thanks guys,  Mark

Mark, if you're using Smaart, the smoothing only affects how the trace is displayed, and so won't affect the outcome of the averaging, which is performed on the raw data. As a proof, take a couple of traces, smooth them down to 1 octave, average them, and then reset the smoothing to NONE on the averaged trace. All the data comes back.

On other platforms, I can't say for sure. I only use Smaart.
Logged

Frank Koenig

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 662
Re: Trace Averaging Separation
« Reply #11 on: March 07, 2019, 10:35:13 pm »

Mark, if you're using Smaart, the smoothing only affects how the trace is displayed, and so won't affect the outcome of the averaging, which is performed on the raw data. As a proof, take a couple of traces, smooth them down to 1 octave, average them, and then reset the smoothing to NONE on the averaged trace. All the data comes back.

On other platforms, I can't say for sure. I only use Smaart.

If the smoothing is done by a linear filter, which typically is the case, and the averaging is the point-by-point mean, then the order in which they're applied doesn't matter, as they are both linear processes. Averaging the smoothed responses will give the same result as smoothing the averaged responses.

I'll reply to Mark's post in more detail tomorrow.

--Frank
Logged
Yes, it is a giant stereo system!

Michael Lawrence

  • Jr. Member
  • **
  • Offline Offline
  • Posts: 72
Re: Trace Averaging Separation
« Reply #12 on: March 07, 2019, 11:16:39 pm »

If the smoothing is done by a linear filter, which typically is the case, and the averaging is the point-by-point mean, then the order in which they're applied doesn't matter, as they are both linear processes. Averaging the smoothed responses will give the same result as smoothing the averaged responses.

I'll reply to Mark's post in more detail tomorrow.

--Frank

What I mean is that smoothing in Smaart specifically is just a graphical thing. The trace averaging operation always uses the original raw measurement data, not the smoothed trace data.
Logged

Mark Wilkinson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 851
Re: Trace Averaging Separation
« Reply #13 on: March 08, 2019, 08:43:52 am »

Mark, if you're using Smaart, the smoothing only affects how the trace is displayed, and so won't affect the outcome of the averaging, which is performed on the raw data. As a proof, take a couple of traces, smooth them down to 1 octave, average them, and then reset the smoothing to NONE on the averaged trace. All the data comes back.

On other platforms, I can't say for sure. I only use Smaart.

Thank you Michael, Yes, I see! Not sure how I've been missing that....

Maybe because I use Smaart for measurements,  then do averaging in FirDesigner.  Still should have seen it though...

In looking at the Smaart user guide this morning, I guess maybe the only way to "fudge" smooth a trace, might be to Export to it  to Ascii and then Import it back in at a lower FFT rate.  But is that even smoothing?  I sure dunno...not going down that path...
Besides, FirD has a number of averaging options I don't even begin to understand, need to learn those first I think...lot's to learn.
As timely luck would have it, I've got class with Jamie next week in Tn.

edit: a little grammar
« Last Edit: March 08, 2019, 11:24:39 am by Mark Wilkinson »
Logged

Mark Wilkinson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 851
Re: Trace Averaging Separation
« Reply #14 on: March 08, 2019, 08:46:18 am »

If the smoothing is done by a linear filter, which typically is the case, and the averaging is the point-by-point mean, then the order in which they're applied doesn't matter, as they are both linear processes. Averaging the smoothed responses will give the same result as smoothing the averaged responses.

I'll reply to Mark's post in more detail tomorrow.

--Frank

Frank, that's great news...sounds like I might be OK despite not knowing what i was doing LOL
thx dude!
Logged

Michael Lawrence

  • Jr. Member
  • **
  • Offline Offline
  • Posts: 72
Re: Trace Averaging Separation
« Reply #15 on: March 08, 2019, 10:43:07 am »

Thank you Michael, Yes, I see! Not sure how I've been missing that....

Maybe because I use Smaart for measurements,  then do averaging in FirDesigner.  Still should have seen it though...

In looking at the Smaart user guide this morning, I guess maybe the only way to "fudge" smooth a trace, might be to Export to it  to Ascii and then Import it back in at a lower FFT rate.  But it that even smoothing?  I sure dunno...not going down that path...
Besides, FirD has a number of averaging options I don't even begin to understand first...lot's to learn.
As timely luck would have it, I've got class with Jamie next week in Tn.

I asked Johnny at Rational and here's what he said (emphasis added):
Quote
Correct, Smoothing in Smaart is just a visual function and does not change the underlying data used for other calculations (like spatial averaging). Captured traces are the raw data, and the smoothing is re-applied upon loading them (hence why you’re not “stuck” with whatever smoothing was used when the trace was captured). Smoothing IS used when you Export/Copy to ASCII, cuz people wanted that.

So your export / import trick would work if that's something you're looking to do.
Logged

Frank Koenig

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 662
Re: Trace Averaging Separation
« Reply #16 on: March 08, 2019, 01:11:52 pm »

Interesting thread...with issues that often drive me crazy !

In my self-taught, measure-in-my-driveway/backyard type experience, I've come to a couple of simple preferences regarding averaging.

I try to average magnitude.
But only when there is high coherence, and response variations tend to be generally the same on and off axis.
So to keep it automated, I smooth measurements to taste before averaging, and the correct the average.
I do this driver by driver, assuming needed corrections are all minimum phase.
If corrections don't respond rather fully, I ditch them as non-minimum phase.
Hopefully I get smoother response optimized for a combination of on and off axis.. 

For phase ...no averaging...it seems to me phase is "inevitably to a spot"
So I use the previously corrected drivers' magnitude response, and phase align those responses to one point, at some distance that is close to equidistant to acoustic centers.

It's never measures as perfect as doing a mag and phase alignment to one spot,
but in today's world of easy to make damn near perfect response at a spot,
it seems like doing so may be its own new can of worms.  (at the expense of off axis)

I hope that made sense...please pass on any critique/observations/reservations on this methodology.


Thanks guys,  Mark

Mark,

This all makes sense. I, too, am a driveway/backyard measurer, although I've recently taken gear out to the Stanford stadium's overflow parking to get more reflection-free space.*

As you point out, aligning phase over multiple measurements is tricky, but only for the excess phase. The minimum phase can be obtained from the log-magnitude and is not subject to variation in the measurement distance. To get the total (and excess) phase across multiple measurements I align the impulse response (IR) peaks. I know this is not theoretically pure but after much experimentation appears to give the most consistent results.
 
When measuring multiple pass-bands I always measure both without moving the mic so that I have accurate phase difference for crossover work. To average I align the HF IR peaks and apply the same time shift to each corresponding LF IR, preserving the phase difference. I set a fixed time window before the HF peak that is large enough to swallow the leading part of the LF IR as well.

To the extent possible, I work with group delay rather than phase to avoid the problems of phase unwrapping and phase origin. Further, the group delays from multiple measurements are straight-forward to average, just like log-magnitudes. The absolute delay of each measurement is irrelevant to the average just as the absolute level of each log-magnitude is irrelevant. The group delay can be obtained directly from the IR by taking the real part of the DFT of the IR multiplied by time and divided by the spectrum. (Nice trick, huh? Plagiarized from the ARTA users' manual.)

To view the phase you integrate the group delay and get to choose the constant of integration which determines the phase origin. I've taken to using the phase directly from the spectrum at some low, fixed frequency, where the phase is well behaved, as the phase origin. This is an "active research area", in other words I don't know what I'm doing and may decide it's all BS.

I use constant proportional-bandwidth Gaussian smoothing, although any reasonable smoothing kernel will give more-or-less the same results. If all you're doing is averaging then you only have to smooth once and it might as well be after averaging. But since I've been playing with non-linear statistical combinations, such as the trimmed mean, I pre-smooth the log-magnitudes to, say 1/24 octave, before combining them. If nothing else, it make it easier to see what 's going on. After combining I smooth to the desired filter resolution, usually 1/6 octave. This also removes any kinks resulting from the median filtering of the trimmed mean. Excess group delays are pretty smooth to begin with so smoothing is less critical, if necessary at all.

--Frank

*The great thing about research universities like Stanford is that people are totally used to odd stuff going on. The place I do my measurements is near a busy bicycle/pedestrian path through the eucalyptus woods. So here's this weird old guy with a van (this alone should raise suspicions ::) ) hoisting big speakers into the air and blasting the heavens with pink noise. And nobody gives it a second look.
« Last Edit: March 08, 2019, 01:16:19 pm by Frank Koenig »
Logged
Yes, it is a giant stereo system!

Frank Koenig

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 662
Re: Trace Averaging Separation
« Reply #17 on: March 08, 2019, 04:19:01 pm »

Another thought on the original topic of spatial sampling: It might be that the required spatial sampling rate depends on the frequency resolution desired in the final result. The more smoothing we're going to apply in the frequency domain the more disparity we can tolerate between spatial samples. This is a problem for folks a lot better at statistics/stochastic signal processing than I am. But I'd like to crack that nut. Thoughts? --Frank
Logged
Yes, it is a giant stereo system!

Mark Wilkinson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 851
Re: Trace Averaging Separation
« Reply #18 on: March 08, 2019, 05:01:26 pm »

Mark,

This all makes sense. I, too, am a driveway/backyard measurer, although I've recently taken gear out to the Stanford stadium's overflow parking to get more reflection-free space.*

As you point out, aligning phase over multiple measurements is tricky, but only for the excess phase. The minimum phase can be obtained from the log-magnitude and is not subject to variation in the measurement distance. To get the total (and excess) phase across multiple measurements I align the impulse response (IR) peaks. I know this is not theoretically pure but after much experimentation appears to give the most consistent results.
 
When measuring multiple pass-bands I always measure both without moving the mic so that I have accurate phase difference for crossover work. To average I align the HF IR peaks and apply the same time shift to each corresponding LF IR, preserving the phase difference. I set a fixed time window before the HF peak that is large enough to swallow the leading part of the LF IR as well.

To the extent possible, I work with group delay rather than phase to avoid the problems of phase unwrapping and phase origin. Further, the group delays from multiple measurements are straight-forward to average, just like log-magnitudes. The absolute delay of each measurement is irrelevant to the average just as the absolute level of each log-magnitude is irrelevant. The group delay can be obtained directly from the IR by taking the real part of the DFT of the IR multiplied by time and divided by the spectrum. (Nice trick, huh? Plagiarized from the ARTA users' manual.)

To view the phase you integrate the group delay and get to choose the constant of integration which determines the phase origin. I've taken to using the phase directly from the spectrum at some low, fixed frequency, where the phase is well behaved, as the phase origin. This is an "active research area", in other words I don't know what I'm doing and may decide it's all BS.

I use constant proportional-bandwidth Gaussian smoothing, although any reasonable smoothing kernel will give more-or-less the same results. If all you're doing is averaging then you only have to smooth once and it might as well be after averaging. But since I've been playing with non-linear statistical combinations, such as the trimmed mean, I pre-smooth the log-magnitudes to, say 1/24 octave, before combining them. If nothing else, it make it easier to see what 's going on. After combining I smooth to the desired filter resolution, usually 1/6 octave. This also removes any kinks resulting from the median filtering of the trimmed mean. Excess group delays are pretty smooth to begin with so smoothing is less critical, if necessary at all.

--Frank

*The great thing about research universities like Stanford is that people are totally used to odd stuff going on. The place I do my measurements is near a busy bicycle/pedestrian path through the eucalyptus woods. So here's this weird old guy with a van (this alone should raise suspicions ::) ) hoisting big speakers into the air and blasting the heavens with pink noise. And nobody gives it a second look.

Too funny Frank......'and nobody gives a second look'  ;D

Your procedure makes sense to me also. I think I understand what you're saying about how you handle minimum and excess phase group delay.
I'll try to give a little depth to my procedure, which greatly mirrors yours.

I see you use ARTA. Does it have an auto-find IR peak, or do you manually set it? 
I ask because I just use Smaart's delay finder to set to peak....when there's no LPF in place.  Even for a sub driver, this has been giving consistent location, with excess phase properly removed.  Then (i think) it's only a matter of how many traces do i want to average.

After I have an average trace I want to work with, I use a FIR file to make all the minimum phase corrections, linear phase crossovers, and phase-only driver corrections if helpful. 
This makes phase alignment of drivers together nearly brain dead easy, because there is no phase slope to deal with as part of total group delay.  It also allows me to I tinker with moving xover frequency up or down, without changing timing. (but must build new FIR files)
And it allows me to nail sub timing almost 'to the sample', even with it's low pass filter in place.....
.....Because I know the post xover group delay has to be the sum of the raw group delay with no LPF in place, and the delay added by the number of samples to impulse centering in the FIR filter. Hope that made sense.
After doing that with lin phase crossovers, I can substitute complementary  IIR crossovers, to knock latency down for live sound, .......and know alignment is spot on despite the usual difficulty of locating a driver with insufficient HF energy.

I'm not as fluent when it comes to windowing or the statistical stuff you speak of...
Sorry i can't add anyhing here...

Good stuff, thx again

edit: fixed some wrong terminology, and backwards HPF and LPF

« Last Edit: March 09, 2019, 08:39:25 am by Mark Wilkinson »
Logged

Russell Ault

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 334
  • Edmonton, AB
Re: Trace Averaging Separation
« Reply #19 on: March 09, 2019, 08:27:25 am »

Another thought on the original topic of spatial sampling: It might be that the required spatial sampling rate depends on the frequency resolution desired in the final result. The more smoothing we're going to apply in the frequency domain the more disparity we can tolerate between spatial samples. This is a problem for folks a lot better at statistics/stochastic signal processing than I am. But I'd like to crack that nut. Thoughts? --Frank

I'm fascinated by all of this; I'm pretty sure your use of spatial averaging and mine are for slightly different purposes.

I'm measuring how speakers operate in a room, and I use spatial averaging to improve the signal-to-noise ratio of what I would like to be a single measurement. I pick a position and measure it, and then take a couple more measurements in close proximity in the hope that those secondary measurements will exhibit different room-bounce characteristics, then take a coherence-weighted average of the lot of them. Comb filtering from room bounce exhibits very low coherence, so that data goes away in the average, and what I'm left with it a trace that looks a heck of a lot like the original measurement, just with much higher coherence and a lot less point-specific room bounce. In this application I still use very little smoothing, because the average is still trying to determine what the sound system is doing at a single point, just with much higher coherence than if I were to only take a single measurement.

-Russ
Logged
Pages: 1 [2] 3  All   Go Up
 


Page created in 0.054 seconds with 23 queries.