Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[mwillems]

I'll confess I'm at a bit of a loss as I don't ordinarily use ASIO to measure (although I use it for playback). ASIO has obvious advantages for playback, but I've never found an advantage to using ASIO for measurements, and there are some significant downsides (As you're discovering).  

Have you tried not using ASIO in any stage of the recording or playback path, just for measurements?  If you configured JRiver to output via WASAPI temporaily you could completely skirt the multiclient ASIO issue, and it should have no real effect on most types of measurement (WASAPI is bitperfect just like ASIO, so there should be no risk that frequency response measurements would differ).  

[packux]

Hi again,

yes indeed using ASIO (or even DirectSound) with Lynx for recording when JRiver is using it with ASIO is next to impossible (or at least i didn't find any meaningfull way of doing it).

So, I indeed proceeded as you also propose, i.e. set-up a  new zone that I use especially for measurements. This zone is using WASAPI to access Lynx and thus I am able to also record using Holms.

There are two lessons learned regarding JRiver through this proccess as well

1. The ASIO Line-In could be better. In its current state, I personally cannot use it even for simpler use cases, such as playing vinyl captured from Lio-8 via JRiver's engine

2. Holms does something funny with JRiver WDM. It seems that after using the WDM with Holms, I either need to completely restart windows, change the speaker configuration in Windows control panel or restart the driver itself to have any audio coming out again from WDM (and JRiver)

[mattkhan]

Close-micing is just what it sounds like: placing the mic very close to the speaker when measuring.  Because the speaker's response will tend to dominate the room response at short distances, you can get a quasi-isolated speaker response by taking measurements close to the speaker.  There's an art to this, and I'll go into it below.  

I wasn't sure whether to post this in this thread or start a new one, it's really about the measurement technique though so I thought I'd post it here as it might be relevant to others trying this.

Context is that I'm starting down the road of some DIY speakers, drivers are on order so I thought I'd try and get my measurement technique sorted using my existing speakers as a test bed. I've struggled to get good results using the techniques listed in this thread before as the speaker is sort of a 2.5 way with 3 tweeters and 2 woofers which makes close mic v hard. I also wanted to consider the impact of DSP on the off axis response. Therefore I set up a stand such that I had at least 1.2m in each direction from the speaker and the mic & I then measured at 7.5 degree intervals and set the windows to remove content after the 1st reflection.

One of the IRs with the window shown, the green arrow is what seems to be the 1st reflection.



and the FR at 0, 7.5, 15 and 22.5 degrees from 200Hz up (about the resolution provided by the window) is in speaker_fr.jpg



The main question that springs to mind here, apart from whether my measurement technique is solid, is how one might decide to handle that kind of response when EQing. It seems the on axis measurement has a bad dip at the XO (1.5kHz is apparently the crossover here) but that dip pretty much disappears as you move off axis. Correct the dip and you get a boost off axis, don't correct the dip and you get a dip on axis. Any useful strategies here? My 1st take is to average those 4 responses and correct based on that. This seems to make the various responses more consistent with each other without any of them being bang on.

[mwillems]

The main question that springs to mind here, apart from whether my measurement technique is solid, is how one might decide to handle that kind of response when EQing. It seems the on axis measurement has a bad dip at the XO (1.5kHz is apparently the crossover here) but that dip pretty much disappears as you move off axis. Correct the dip and you get a boost off axis, don't correct the dip and you get a dip on axis. Any useful strategies here? My 1st take is to average those 4 responses and correct based on that. This seems to make the various responses more consistent with each other without any of them being bang on.

Several thoughts:  

1) What does the phase look like in the crossover region where the dip is?  That would tell you definitively whether the issue was crossover based or not.  If it is crossover based, you could potentially correct the phase cancellation issue (using convolution), which would have the secondary effect of fixing the on axis response without necessarily unduly prejudicing the off-axis response.

2) I have found gating to be unreliable.  Have you taken any measurements in Holm (or exported to Holm) to take advantage of Frequency Dependent Windowing? Gating before first reflection (in my own experience) does not always produce repeatable results or successfully approximate close micing. Because FDW produces consistent, comparable, repeatable results for me, I would recommend using it whenever possible instead of a gating system.  

3) If you measure at one foot, is the dip still present?  How about at two meters?

4) It looks to me (although it's a little hard to disentangle) that both the on axis and the 7.5 degree graphs look like they have a multi-dB dip at 1.5KHz.  The 15 and 22 degree still have a 1dB dip there.  Am I misreading that?   I think (if it were not a phase issue but an actual frequency dip), you could safely add at least 1dB there without any trouble at all.  How far off axis are most of your seats?  

5) Averaging the responses creates some risks because your off axis response is pretty different, and not just at 1.5KHz (The multi-dB dip off-axis at 2.5KHz might skew things a bit).  Unless all of your seats are significantly off axis, I'd recommend trying for the flattest on axis response possible (recognizing that you can't directly correct some crossover issues).  Off-axis falloff isn't always linear (i.e. it isn't necessarily dB for dB depending on the polar pattern) so you might find that the off-axis doesn't change as much as you expect when you correct the on-axis.

All that said, my first recommendation in re: the 1.5Khz notch is to have a look at the phase.  If it's obvious that you're getting phase related cancellation on axis, then try correcting the phase (if you can use convolution in your application) and see if it helps the FR.  

If that's not it, I'd recommend either retaking the measurements in Holm or just exporting the impulses to Holm, and looking at them with FDW to see if the dip is the same, different, or what.  I know that it "shouldn't matter for frequencies above the gate," but that has not been my own experience at all.  For example, I can take two measurements from the same distance, but with a slightly different mic position (I'm talking inches), and the two measurements (when gated) will look very different above the gate.  This is not at all the case with FDW.  All of my FDW'ed 1 meter measurements look more or less the same (give or take minor fluctuations) even if the mic is in a slightly different position, and they look at least comparable in shape to my FDW'd 1 foot measurements.  

It's well documented that, at minimum, gating has a strong smoothing effect for an octave or two above the "gate frequency," which you can easily see if you adjust the gate on an existing measurement.  So while the data "below the gate" is guaranteed to be unreliable, even the data above the gate can be a little flaky.  For these reasons, I'm not convinced that gating actually does a very good job of simulating close micing or separating out the room, probably because late reflections aren't all there is to the story.  

I'm sure that sounds a little categorical, but I can't tell you how many hours I spent trying to get good results with gating, and getting nowhere or working in circles because my measurements never looked the same from day to day.  I finally lost my patience with it when my measurements on a Sunday showed that all of the EQ (every filter) that I'd spent an hour dialing in on Saturday was unnecessary/counterproductive.  After making the switch to FDW, I've never looked back.

P.S.- I know you said you had difficulty close micing due to multiple drivers. Did you have trouble getting consistent measurements, or was it just obvious that the very close measurements were incorrect/had problems?  I'm sort of curious where the process "broke down" for you and whether you might not be able to measure profitably closer than 1 meter?

[natehansen66]

Matt - change your dB scale to show a 40dB range....that will give you 5dB divisions which is a standard way to display data. Your 1dB scale can make things look much worse than they really are. What kind of speakers are we looking at here?

mwillems - when you talk about having repeatability issues with gating are you talking about measurements for room correction from the lp?

It sounds like matt is looking to get quasi-anechoic data on his speakers and in my experience as long as you are in the far-field then repeatability is not an issue with gating. I've measured my speakers many times on different days and the result was within 1dB. Now if you're talking about different results with gated data from the lp then that I can understand for the reasons mwillems has given. I  agree that quasi-anechoic data taken in a typical living room is suspect below about 1khz, but I'm not sure that a close mic measurement is sufficient to capture baffle diffraction effects.

[mwillems]

mwillems - when you talk about having repeatability issues with gating are you talking about measurements for room correction from the lp?

I'm talking about from any room position farther than about 1 meter out to the listening position; basically anything other than close micing (i.e. out in the room) produces inconsistent results for me with static gating.

I  agree that quasi-anechoic data taken in a typical living room is suspect below about 1khz, but I'm not sure that a close mic measurement is sufficient to capture baffle diffraction effects.

I think we mostly agree, with the modification that I think that gated data in a living room is suspect below about 1KHz (or higher depending on the gate).  FRD applies a variable gate that scales to the wavelength, so it will produce useful data at all wavelengths.  That, in part, is why I think it produces more consistent results (at least for me).  I can get a very similar 300Hz measurement on axis at 3 feet and at 10 feet with FRD; with a static gate, it's not at all similar (and as I said above, sometimes differs for me in a similar position on different days).

And I definitely agree that close micing (really close micing) will not capture baffle diffraction. Some farther field measurement (at least at a meter out) is desirable, but my experience with gated measurements is that I get much less consistent results (especially below 1K, but also in general) with a static gate than with frequency dependent windowing.  

It sounds like you've had better luck with static gating than I have, maybe, if you're willing, could you can offer some pointers on gate configuration?

[mattkhan]

@nate the speaker in question is an MK MP150 Mk2 which has 2 5.25" woofers and 3 1" dome tweeters. AIUI the woofers cross to 2 of the tweeters at 1.5kHz and then those the 3rd tweeter joins in somewhere around 4kHz.

There are further measurements of that speaker in http://yabb.jriver.com/interact/index.php?topic=88942.msg614015#msg614015 and some discussion with @mwillems ensues the problem with measurements taken closer in was, I think, down to the no of drivers involved making it hard to get a clear picture of what was going on. The problem with moving to 2m is that, in my room, I'm practically at the listening position at that point. I didn't have the relevant kit to get the speaker off the wall and measure it in isolation at that point though, I do now (hence these new measurements) so I am planning to get it outside and measure which would give me ~3m clear space (apart from the floor so was thinking of laying down some material to absorb between mic and driver).

FWIW I've shared the mdat at https://drive.google.com/file/d/0BxdmSMpV-t3GQXNNMC1DWGt3VVU/view?usp=sharing in case there is some Q left unanswered by my point by point comments belows.

1) What does the phase look like in the crossover region where the dip is?  That would tell you definitively whether the issue was crossover based or not.  If it is crossover based, you could potentially correct the phase cancellation issue (using convolution), which would have the secondary effect of fixing the on axis response without necessarily unduly prejudicing the off-axis response.

2) I have found gating to be unreliable.  Have you taken any measurements in Holm (or exported to Holm) to take advantage of Frequency Dependent Windowing? Gating before first reflection (in my own experience) does not always produce repeatable results or successfully approximate close micing. Because FDW produces consistent, comparable, repeatable results for me, I would recommend using it whenever possible instead of a gating system.  

I use acourate which does have a FDW function, the Q is how to apply it. One user described the controls in http://digitalroomcorrection.hk/http___www.digitalroomcorrection.hk_/Driver_Linearization.html which says

FDW means a frequency dependent windowing defined in terms of cycles.
A 1kHz sine wave has a cycle time of 1 ms. So a FDW value of 15 means a window width of 15 ms.
With a 20 Hz sine wave the same FDW value means a window width of 15 * 50 ms = 750 ms.

In the FDW panel you find 4 parameters. a/b, c/d
The parameters control the windowing of a given single pulse response. a/b is a half window on the left side of the pulse peak, c/d is a half window on the right side of the pulse peak. In normal cases you can simply set a=c and b=d = symmetric windowing.
a and c define the width at low frequencies, whereas b and d define the width at high frequency (fs/2). In between the values are interpolated.

For example, here's that on axis measurement with 15/15 15/15 FDW vs 4/4 4/4

3) If you measure at one foot, is the dip still present?  How about at two meters?

see the post I linked to earlier for measurements taken closer albeit they were taken on wall

4) It looks to me (although it's a little hard to disentangle) that both the on axis and the 7.5 degree graphs look like they have a multi-dB dip at 1.5KHz.  The 15 and 22 degree still have a 1dB dip there.  Am I misreading that?   I think (if it were not a phase issue but an actual frequency dip), you could safely add at least 1dB there without any trouble at all.  How far off axis are most of your seats?  

the dip at 1.5kHz is pretty much gone by 22.5 degrees, it gradually softens before that point.

The seating area is covered by

L = 0 to -32 degrees
R = 0 to +32 degrees
C = +/- ~16 degrees

The wall on one side has an open fireplace so is somewhat chaotic/unmanageable for the purposes of first reflections. The other wall (ipsilateral to the L speaker) is a flat surface.

1st reflection points are in the ranges of

L = ~15-40 degrees
C = ~40-50 degrees
R = ~40-50 degrees

to illustrate



I'm planning a SEOS10 + AE TD10H build to replace these.

5) Averaging the responses creates some risks because your off axis response is pretty different, and not just at 1.5KHz (The multi-dB dip off-axis at 2.5KHz might skew things a bit).  Unless all of your seats are significantly off axis, I'd recommend trying for the flattest on axis response possible (recognizing that you can't directly correct some crossover issues).  Off-axis falloff isn't always linear (i.e. it isn't necessarily dB for dB depending on the polar pattern) so you might find that the off-axis doesn't change as much as you expect when you correct the on-axis.

the problem I have atm is that a listening position only based correction is resulting in recessed mids (~600-1000Hz) which isn't ideal. This is what acourate sees as the basis for correction & the inverse that it uses as the basis for the correction filter.



From the top

Red = the in room response after FDW and other smoothing
Brown = a "normal" bk style target curve
Green = the inverse from the bk curve that will form the basis of the filter

Blue = an adjusted bk style curve to accommodate that "step" around 1kHz
Black = the inverse of the adjusted bk curve

You can see in the red trace the relatively high level in that midrange. The normal curve then cuts this down to size to produce a flat response BUT this presents to my ears as recessed mids (voices are somewhat veiled and pushed backwards in the sound stage). I attempted to hack around this with the blue curve which introduces a hard step in the target curve, this does bring the voices forward (as you would expect) but wasn't ideal in some other respects.

I have run through a correction based on an average of the FDW'ed responses, only briefly listened to it though. I'll post that info in another post for reference.

[mattkhan]

(not sure whether we should split this off to a separate topic? perhaps a mod can move if they want to)

This is correction I came up with last night based on those measurements. I used default FDW values in acourate which is 15 cycles before/15 cycles after the peak, arguably too large a window but I wanted to work through the process to see what came out the other end.

This shows the 0/7.5/15/22.5 degree measurements after FDW and then the average of the 4 (offset below)



I decided not to attempt to correct below 1kHz and so this produces the following, red is the corrected high pass and the other traces are the same measurements put through the XO (remember that acourate works by creating a set of XO pulses which you can apply filters to etc, it then convolves the sweep with the XO and routes that to each channel for measurement purposes, you then do room correction against the resulting measurement and the final filters are a combination of that correction + the XO)



This produces the following measurement at the listening position, green is with a vanilla XO and red is with this corrected XO



There is a bit of an ugly spike at 4kHz & the upper frequency droop is a concern but the response is much more consistent from 300Hz to 4kHz.

Finally the measured results of correction, red is the vanilla XO and green is with the new corrected XO.



Not radically different tbh, somewhat smoother through the 800-2kHz range though.

[mattkhan]

getting somewhat back on topic.... is data with this resolution (i.e. maybe 5ms of clean data) sufficient for passive XO design when the XO is likely to be in the 1.5kHz range? I don't have a bigger space in the house so would have to head outdoors to get better data.

[mwillems]

Quite a bit to respond to, and now that you linked I recall some of our prior discussions (apologies, it's been an eventful year for me  )

That phase makes it relatively clear that the issue is phase cancellation related and the crossover's delay isn't quite right.  Looking at your correction in acourate, it's obvious that some phase correction is happening in tandem with the frequency response correction (or it wouldn't be able to fill the hole at all).  But it's still looks like it's adding undesired boost off-axis.  All told the projected correction is plus or minus 1.5 dB so it may not matter too much.  If you want to try and squeeze the problem down to nothing, here's what I'd suggest investigating: Is there any way in acourate to dial in specific phase changes without also changing frequency?   The reason I ask is that is that RePhase has a feature that allows to basically just reverse the phase shift of a passive crossover, which can resolve some phase cancellation issues without further monkeying with FR.  If you do that you might be able to get things normalized on axis with less mischief off axis.

For example, here's that on axis measurement with 15/15 15/15 FDW vs 4/4 4/4

Thanks for humoring me on this, this is exactly what I was looking for.  In the guide above I recommend using a coefficient of 12 for FDW in Holm, but I'm not sure how Holm's coefficients line up with Acourate's.  Somewhere between Acourate's 4 and 15 to be sure.

see the post I linked to earlier for measurements taken closer albeit they were taken on wall
the dip at 1.5kHz is pretty much gone by 22.5 degrees, it gradually softens before that point.

The seating area is covered by

L = 0 to -32 degrees
R = 0 to +32 degrees
C = +/- ~16 degrees

Well that makes it all easier (and thanks for the excellent drawings), because each seating position appears to be on axis (ish) for one speaker while being somewhat off axis for the other.  That's great when you have complementary on and off-axis peaks and dips because the on-axis speaker will "compensate" for the off-axis speaker.  So if you're up 1.5 dB off axis, but down 1.5 dB on axis it will present relatively coherently at all of your listening positions, if you see what I mean.  The fireplace is obviously a wildcard, but you get the idea.

It looks to me, though, that if you really wanted to "solve" the 1.5 KHz problem that manual phase manipulation might be the way to do it.

You can see in the red trace the relatively high level in that midrange. The normal curve then cuts this down to size to produce a flat response BUT this presents to my ears as recessed mids (voices are somewhat veiled and pushed backwards in the sound stage). I attempted to hack around this with the blue curve which introduces a hard step in the target curve, this does bring the voices forward (as you would expect) but wasn't ideal in some other respects.

That's very odd; It's obvious based on your final measurements below that Acourate is overcorrecting in the midrange (600 to 1000); any ideas what combination of factors is leading it to see a phantom peak there?  Do you see that same 4dB 600 to 1000 peak at the listening position in other software like REW or Holm?

This produces the following measurement at the listening position, green is with a vanilla XO and red is with this corrected XO


There is a bit of an ugly spike at 4kHz & the upper frequency droop is a concern but the response is much more consistent from 300Hz to 4kHz.

Finally the measured results of correction, red is the vanilla XO and green is with the new corrected XO.



Not radically different tbh, somewhat smoother through the 800-2kHz range though.

I actually think that's a pretty dramatic improvement (given that you're down to dB by dB variations).  You managed to iron out a 2dB dip (at 1.2K), a 1.5dB dip at 600Hz, and flattened a +/- 1dB swing (between 750 and 900) in the region of interest. I'd call that a pretty serious victory, especially since the dips/ripples in question were relatively high Q dips (Which are more likely to be audible).  It's still sort of a puzzle why Acourate was overcorrecting in the first place.

getting somewhat back on topic.... is data with this resolution (i.e. maybe 5ms of clean data) sufficient for passive XO design when the XO is likely to be in the 1.5kHz range? I don't have a bigger space in the house so would have to head outdoors to get better data.

Absolutely, especially if you'll be able to close mic the drivers you're working with.  "Most of my design has benn for active crossovers, but I've assisted on one or two passives too.  

Active crossover design is, of course, easier because you can change it after implementing (sometimes on the fly), but (with that caveat in mind) I've designed a few active crossovers with the speakers in a normal living room (one with a cross at 800Hz, and one with a cross at 350Hz), and they worked out great. The crossovers worked correctly when the speakers were moved to other environments, etc.  

You can also use JRiver to "demo" a passive design by dialing it in as a n active filter and measuring.  So you never need to rely on simulations, you can test the crossover design and adjust as necessary without ever soldering a part.  You can basically use JRiver as a live crossover design suite that can be tweaked on the fly.  This has saved me some serious grief as I'm not very handy with a soldering iron.  By far the hardest part of any crossover design is setting the interchannel delay correctly, which is trivial when you can just change the delay at will like you can in JRiver.

As long as you can close mic and have software that provides for frequency dependent windowing, you can accomplish what you need for basic crossover design purposes; if you can demo them as an active crossover first, you can get them close to perfect.

Out of curiosity, any particular reason for going with a passive crossover design?  There are some real challenges that go along with passive crossovers (lost efficiency, lack of slope flexibility, correct delay i very challenging to accomplish passively, slopes change due to changes in temperature and the aging of components, etc.); there are obviously significant pros too (convenience, only need a stereo amp, easier for end users, etc.), but have you considered doing an active/all software crossover?  Sorry to be meddlesome if you've already considered and rejected the idea, but you really can do much more with active correction if you're starting from scratch in terms of building speakers; if you're interested I'm happy to provide some links outlining the benefits in more detail.

[mattkhan]

Thanks for the detailed response.

 If you want to try and squeeze the problem down to nothing, here's what I'd suggest investigating: Is there any way in acourate to dial in specific phase changes without also changing frequency?   The reason I ask is that is that RePhase has a feature that allows to basically just reverse the phase shift of a passive crossover, which can resolve some phase cancellation issues without further monkeying with FR.  If you do that you might be able to get things normalized on axis with less mischief off axis.

I don't know tbh. Theoretically it seems like I should be able to generate a filter externally and then convolve the acourate XO with that filter before using it in the sweep. On the other hand, acourate has loads of features I don't remotely understand so perhaps it can be done in there instead. Can you describe how you would do this in RePhase?

 Thanks for humoring me on this, this is exactly what I was looking for.  In the guide above I recommend using a coefficient of 12 for FDW in Holm, but I'm not sure how Holm's coefficients line up with Acourate's.  Somewhere between Acourate's 4 and 15 to be sure.

I imported the impulse into holm and compared the two visually, I think it's actually more like 18-20 in acourate. Holm also indicates that the window is a symmetric window & is 20ms wide on the left and right at 1kHz. The shape looks like a Hann window but I'm not 100% sure about that. I think it's reasonable to argue that 12 in Holm is too big in this case given that I know the HF reflection arrives after 5ms or so.

 
That's very odd; It's obvious based on your final measurements below that Acourate is overcorrecting in the midrange (600 to 1000); any ideas what combination of factors is leading it to see a phantom peak there?  Do you see that same 4dB 600 to 1000 peak at the listening position in other software like REW or Holm?

No I don't know, acourate is a quite high maintenance partner at times I think probably 1 in 4 times I measure I get some sort of anomaly.

 
I actually think that's a pretty dramatic improvement (given that you're down to dB by dB variations).  You managed to iron out a 2dB dip (at 1.2K), a 1.5dB dip at 600Hz, and flattened a +/- 1dB swing (between 750 and 900) in the region of interest. I'd call that a pretty serious victory, especially since the dips/ripples in question were relatively high Q dips (Which are more likely to be audible).  It's still sort of a puzzle why Acourate was overcorrecting in the first place.

I have completed some initial listening and it does seem improved. I do find myself getting more sensitive to these variations as time goes by, possibly not a good thing!

 
You can also use JRiver to "demo" a passive design by dialing it in as a n active filter and measuring.  So you never need to rely on simulations, you can test the crossover design and adjust as necessary without ever soldering a part.  You can basically use JRiver as a live crossover design suite that can be tweaked on the fly.  This has saved me some serious grief as I'm not very handy with a soldering iron.  By far the hardest part of any crossover design is setting the interchannel delay correctly, which is trivial when you can just change the delay at will like you can in JRiver.

As long as you can close mic and have software that provides for frequency dependent windowing, you can accomplish what you need for basic crossover design purposes; if you can demo them as an active crossover first, you can get them close to perfect.

this sounds v promising. How do you translate the passive design into an active filter? I had noticed soundeasy had this capability but, from what I've read, it is rather pernickity about hardware and is a bear to learn.

 
Out of curiosity, any particular reason for going with a passive crossover design?  There are some real challenges that go along with passive crossovers (lost efficiency, lack of slope flexibility, correct delay i very challenging to accomplish passively, slopes change due to changes in temperature and the aging of components, etc.); there are obviously significant pros too (convenience, only need a stereo amp, easier for end users, etc.), but have you considered doing an active/all software crossover?  Sorry to be meddlesome if you've already considered and rejected the idea, but you really can do much more with active correction if you're starting from scratch in terms of building speakers; if you're interested I'm happy to provide some links outlining the benefits in more detail.

The main reason is that I have cables chased into a solid brick wall (my equipment rack is some distance away under the stairs) and running a 2nd pair is invasive work. The secondary reason is that I intend to remediate this in a few years when we have an extension built so I figured I would go passive for now and swap to active later. Finally, and slightly perversely, passive xo design seems like an interesting challenge.

Certainly interested in any and all links and/or adivce you have anyway. I suspect I'm going to need it

[mwillems]

Thanks for the detailed response.
I don't know tbh. Theoretically it seems like I should be able to generate a filter externally and then convolve the acourate XO with that filter before using it in the sweep. On the other hand, acourate has loads of features I don't remotely understand so perhaps it can be done in there instead. Can you describe how you would do this in RePhase?

There is a tab in rephase labelled "filters linearization"; it includes a column labelled crossovers which includes a large number of phase manipulation presets that effectively just reverse the phase wrap introduced by many conventional crossover filters without touching the FR.

I imported the impulse into holm and compared the two visually, I think it's actually more like 18-20 in acourate. Holm also indicates that the window is a symmetric window & is 20ms wide on the left and right at 1kHz. The shape looks like a Hann window but I'm not 100% sure about that. I think it's reasonable to argue that 12 in Holm is too big in this case given that I know the HF reflection arrives after 5ms or so.

That makes sense; I'll make a mental note of that for future comparability.

this sounds v promising. How do you translate the passive design into an active filter? I had noticed soundeasy had this capability but, from what I've read, it is rather pernickity about hardware and is a bear to learn.

The short answer is that I don't, I do the reverse.  I'm not an EE and have limited skill with practical electronics, so I do all of my initial design empirically.  I'm not good enough at circuit analysis to "work backwards" from an existing circuit layout to a filter bank unless the filters involved are textbook filters.

So I just work it out empirically; I measure the speaker elements separately, then figure out what filters and delay I need, dial them into JRiver, re-measure, etc. until I figure out the ideal combination.  Once you know the actual filters you need, it's a matter of finding the right parts and layout, which is obviously non-trivial, but most of the PEQ filters in JRiver are textbook filters. So you can use the various calculators and design tools floating around the internet, like this one that will spit out a symmetrical crossover circuit complete with parts values: http://www.diyaudioandvideo.com/Calculator/XOver/  

Once you know what the ideal is, the real issue at that point is reducing complexity and trying to figure out to how to do the most with the fewest filters.  I'll confess that if I were making a passive filter tomorrow, I'd want to huddle up with an EE friend of mine to make the translation from filters to an actual circuit to keep myself out of trouble.   I tried doing one on my own many years back and the results were suboptimal  

Delay will be a problem no matter what; doing fine delay in analog is brutally hard, which is why most speaker designers solve the problem mechanically instead: they just skip the issue of electronic delay entirely and try to get the drivers' voice coils physically aligned so delay isn't so important.  There are obviously limits to that method, especially if you have horns, etc.

The main reason is that I have cables chased into a solid brick wall (my equipment rack is some distance away under the stairs) and running a 2nd pair is invasive work. The secondary reason is that I intend to remediate this in a few years when we have an extension built so I figured I would go passive for now and swap to active later. Finally, and slightly perversely, passive xo design seems like an interesting challenge.

Certainly interested in any and all links and/or adivce you have anyway. I suspect I'm going to need it

Fair enough; I enjoy a good challenge myself  .  I'm happy to help to the extent I can, with the caveat that (as noted above) I'm pretty iffy at circuit design.

When you reach the point that you're thinking about active design, the first article I'd suggest is this famous one by Rod Elliot: http://sound.westhost.com/bi-amp.htm

It does a very good job of laying out some of the major advantages of active crossovers and bi-amping, and contains links to a number of very good articles about crossover design.

To "amplify" and offer some observations of my own:

1) Delay is very challenging to get right in a passive crossover so your phase coherence will always be approximate (unless you're a very good EE or can get cm-tolerance driver alignment)
2) Because of the size and cost of electronic components (as well as power losses from passive components), you'll be limited in the amount of filtering that can be done with a passive crossover.  For example, 4th Order crossover slopes are both parts-intensive and extremely power hungry, which makes them less practical in passive crossovers.  Ditto for other types of parametric filtering.
3)  This isn't the end of the world because we can dial in an arbitrary amount of electronic DSP in software, but software DSP will be very limited in it's ability to correct any FR anomalies in the crossover region because both speakers won't be contributing to them equally, so you'll have some unresolvable lumps and dips in the crossover region.  This is exacerbated because of the design limitations in 2) lead to most passive crossovers using lower-order crossover slopes (1st, 2nd, or 3rd), which means the effective "crossover region" could easily be three to five octaves!
4) Active crossovers with bi-amping significantly reduce intermodulation distortion
5) Because of the power losses mentioned in 2) and because passive crossovers require padding down the more sensitive of the two drivers, active crossovers allow for significantly better power efficiency.
6) Passive components change with temperature and age (i.e. the capacitance of a capacitor is not the same at all temperatures or after a few years).  Because capacitance/inductance/etc. determine the frequencies/slopes of the filters, you may find that you get different speaker performance when playing music loud (because the parts heat up due to power flow, etc.).  This can be mitigated to some extent by buying nicer components that have better temperature tolerance, but that raises cost, etc.

Obviously none of these issues is insurmountable, and there are many very high quality passively crossed speakers out there.  It's just that passive crossovers necessarily involve making some compromises, especially for a DIYer. They can be close to perfect if you're a talented designer with a wide-open budget, but if you're trying to piece it together (or do it at low cost) it can be harder to get a good result.  

By contrast, if you have the equipment to do active crossovers/bi-amping (a few extra amp channels and a multi-channel DAC) you can skip those compromises entirely, and just hack away until it's actually right (or let Acourate do it for you)  

[mattkhan]

There is a tab in rephase labelled "filters linearization"; it includes a column labelled crossovers which includes a large number of phase manipulation presets that effectively just reverse the phase wrap introduced by many conventional crossover filters without touching the FR.

OK I see it, I will have a play around with that. I think acourate has the same functionality with its excess phase correction but I'm not sure how to drive it in this situation.

The short answer is that I don't, I do the reverse.  I'm not an EE and have limited skill with practical electronics, so I do all of my initial design empirically.  I'm not good enough at circuit analysis to "work backwards" from an existing circuit layout to a filter bank unless the filters involved are textbook filters.

So I just work it out empirically; I measure the speaker elements separately, then figure out what filters and delay I need, dial them into JRiver, re-measure, etc. until I figure out the ideal combination.  Once you know the actual filters you need, it's a matter of finding the right parts and layout, which is obviously non-trivial, but most of the PEQ filters in JRiver are textbook filters. So you can use the various calculators and design tools floating around the internet, like this one that will spit out a symmetrical crossover circuit complete with parts values: http://www.diyaudioandvideo.com/Calculator/XOver/ 

right ok, I may try that approach. My basic plan so far was to try and get PCD working but we'll see how that goes I have to twiddle my thumbs until the drivers arrive anyway.

[mattkhan]

OK I see it, I will have a play around with that. I think acourate has the same functionality with its excess phase correction but I'm not sure how to drive it in this situation.

FWIW acourate does have this feature and automagically does it for you if you know how. Basically you create an impulse that contains the excess phase only, reverse it and then you can convolve the XO used in the sweep with that excess phase correction. To simulate the effect I convolved my original impulse with the phase correction and then dumped it in REW to show the comparison



apparently this is unlikely to work well in practice but I'll give it a whirl next time I measure anyway just to see

[mwillems]

apparently this is unlikely to work well in practice but I'll give it a whirl next time I measure anyway just to see

Out of curiosity, why is it unlikely to work well?  If you can alter the phase relationships, you can probably sort out at least some of the on axis cancellation (that's been my own experience anyway).  Or is it that it's less likely to work well than going the whole distance with Acourate?  Phase wrap isn't super audible when the two sides are phase coherent, but they probably are not completely phase coherent in your case (or you wouldn't have that dip).

Regardless, I'll be interested to hear whether it helps at all. 

[mattkhan]

Out of curiosity, why is it unlikely to work well?  If you can alter the phase relationships, you can probably sort out at least some of the on axis cancellation (that's been my own experience anyway).  Or is it that it's less likely to work well than going the whole distance with Acourate?  Phase wrap isn't super audible when the two sides are phase coherent, but they probably are not completely phase coherent in your case (or you wouldn't have that dip).

I hope he won't mind me quoting him (Uli B) but he said

You see a dip in the XO transition area because the phases of both drivers are different. But you only have the answer of the sum of both drivers. So of you apply a filter which will remove the dip it will boost both drivers. And when you measure off axis then again the frequency response will change.
The same happens with a phase changing filter. It will equally change the phase of both drivers. So it will correct the phase sum of them. But changing the position to off axis will also change again the sum.

Regardless, I'll be interested to hear whether it helps at all. 

apparently not, does this mean that dip is not phase related at all? brown is before, green is after btw

[mwillems]

That certainly doesn't look phase related in those measurements. The final test: if you add boost there, does it actually fill in?  If you can EQ it one for one, it's definitely not phase cancellation

[GiAnt]

Out of curiosity, any particular reason for going with a passive crossover design?  There are some real challenges that go along with passive crossovers (lost efficiency, lack of slope flexibility, correct delay i very challenging to accomplish passively, slopes change due to changes in temperature and the aging of components, etc.); there are obviously significant pros too (convenience, only need a stereo amp, easier for end users, etc.), but have you considered doing an active/all software crossover?  Sorry to be meddlesome if you've already considered and rejected the idea, but you really can do much more with active correction if you're starting from scratch in terms of building speakers; if you're interested I'm happy to provide some links outlining the benefits in more detail.

Mwillems, the tutorial in the first page of this thread is impressive ... It is very, very instructive. Many thanks! You already help me some days ago in the fine tuning of crossover with JRiver MC. My ultimate goal is to manage with JRiver MC my DIY multiway active system (actually based on external DSP). Unfortunately, I have still limited knowledge on this topic to correctly arrange by myself the whole system. I will be very grateful to you if you could show me the "flowchart" (possibly with some links or pictures) to follow to realize an entirely Pc-based system, from crossovers to room correction. I will use the information to write a guide on the usage of JRiver MC in this context, on an italian forum of digital-Hi-Fi-audio. In order to ensure a large spreading of the tutorial, the usage of free software (such as Holm and Rephase) is wellcome.
Thanking you in advance

[mwillems]

Mwillems, the tutorial in the first page of this thread is impressive ... It is very, very instructive. Many thanks! You already help me some days ago in the fine tuning of crossover with JRiver MC. My ultimate goal is to manage with JRiver MC my DIY multiway active system (actually based on external DSP). Unfortunately, I have still limited knowledge on this topic to correctly arrange by myself the whole system. I will be very grateful to you if you could show me the "flowchart" (possibly with some links or pictures) to follow to realize an entirely Pc-based system, from crossovers to room correction. I will use the information to write a guide on the usage of JRiver MC in this context, on an italian forum of digital-Hi-Fi-audio. In order to ensure a large spreading of the tutorial, the usage of free software (such as Holm and Rephase) is wellcome.
Thanking you in advance

I've been working on a guide to active speakers/crossover design in JRiver, but I have an infant daughter so the guide has languished unfinished on my "to do" pile.  I'll post here when it's finished (if it's ever finished), but in the meanwhile if you have any specific questions I'm happy to answer.  You may want to start another thread so we can get deep into the issues.

For now, I'll share the two pieces of advice I wished I'd known when I started building active speakers.  You may already know this, but if I had known these two things at the start I'd have saved myself some time:

1) The most important thing is to measure your results using a calibrated microphone with logarithmic sweeps using frequency dependent windowing.  You can have things worked out in theory, but your measurements will tell you whether you've done things correctly, and will identify problems or things you may not have thought through all the way.  An uncalibrated mic will fool you, and using pink noise for anything other then phase coherence or setting relative volume levels will fool you, and gated sweeps will also fool you (unless you know how to read them).  Do sweeps with frequency dependent windowing with a calibrated mic, and you'll get where you need to be faster.

2) If you want crossovers that are likely to "just work" without much fuss, use symmetrical 4th order Linkwitz-Riley filters.  They're phase coherent and sum flat.  Almost any other topography (except for 8th order Linkwitz-Riley) will require special care and feeding, which may be desirable once you know exactly what you're doing, but to start with, LR 4 will make your life much easier.  You can make 4th order linkwitz rileys in JRiver's parametric equalizer by creating two identical 12dB lowpasses and two identical 12dB highpasses (stacking two butterworth filters makes a linkwitz riley filter).

[GiAnt]

... I have an infant daughter so the guide has languished unfinished on my "to do" pile. 

This is a very good news. I'm sure your best DIY project!

You may want to start another thread so we can get deep into the issues.

Ok, one of the next days I will start a new thread on this topic

1) The most important thing is to measure your results using a calibrated microphone with logarithmic sweeps using frequency dependent windowing.  You can have things worked out in theory, but your measurements will tell you whether you've done things correctly, and will identify problems or things you may not have thought through all the way.  An uncalibrated mic will fool you, and using pink noise for anything other then phase coherence or setting relative volume levels will fool you, and gated sweeps will also fool you (unless you know how to read them).  Do sweeps with frequency dependent windowing with a calibrated mic, and you'll get where you need to be faster.

Ok, I have the equipment and the softwares to perform this.

2) If you want crossovers that are likely to "just work" without much fuss, use symmetrical 4th order Linkwitz-Riley filters.  They're phase coherent and sum flat.  Almost any other topography (except for 8th order Linkwitz-Riley) will require special care and feeding, which may be desirable once you know exactly what you're doing, but to start with, LR 4 will make your life much easier.  You can make 4th order linkwitz rileys in JRiver's parametric equalizer by creating two identical 12dB lowpasses and two identical 12dB highpasses (stacking two butterworth filters makes a linkwitz riley filter).

Yes, I know. I will follow your advice.
See you

[harlington]

First of all and as many has said, thanks so much for a well written thread. My question is that we are listening to speakers, subs in room, does it make more sense to measure and eq speakers/subs at the main listening position? Is it ok to average the left and right measurements in REW and eq it (the average measurement)? I have a 7.1 system that I use a 100hz crossover to my subs (all 7 speakers), I was playing with the REW eq to eq the center and subs, the average of front left and right plus subs measurements, the average of side L/R surround plus subs and the average of rear L/R surround plus subs (all use full range in REW filter). I only match target in REW eq from 10-20khz for my center plus subs, and from 80-20khz for the rest so that it does not apply filters to subs many times. As the result, I have 4 set of filters (center, main, side, and rear) that I will manually enter in jriver DSP. In jriver, I currently have two  separate parametric eqs that I plan to type in the eq filters for my center and main (38 total filters) using jriver parametric eq and type in filters for my side and rear surrounds using parametric eq2 (10 total filters). Is that ok? Is there a max number of filters one can enter in jriver DSP? Can I enter all 48 filters in one parametric eq? Also, will the change take effect immediately by ticking and unticking the parametric eq?  Thanks for all the help as I am new to jriver and REW eq.

[mwillems]

There's a lot of information in your post and I'm having trouble following some parts of your setup.  I'll answer some of the questions that "pop out," but if I miss something, let me know  

First of all and as many has said, thanks so much for a well written thread. My question is that we are listening to speakers, subs in room, does it make more sense to measure and eq speakers/subs at the main listening position?

I don't recommend measuring at the listening position for correction unless you plan on listening only in that one small position or unless you're using one of the more sophisticated commercial automatic measuring suites (like audiolense and acourate).  If you re-read the opening of the guide, I went into some detail (with measurements) to show why I think measuring at the listening position is not the best idea.  The bottom line is that speaker non-linearities can often be usefully EQ'd, and certain kinds of near-field room effects (like corner gain) can also be meaningfully eq'd.  Other mid and far field room effects (rear wall bounce, nulls, nodes and anti-nodes, etc.) cannot be meaningfully eq'd so measuring them and trying to correct them is not frutiful.  

Additionally, correcting for your listening position will not necessarily produce good results away from your listening position (even as close as a foot or two).  Have a look at my graphs above for a demonstration of that.  My advice is to follow the guide above, and then try to selectively improve things at your listening position from there (checking to make sure it's working as intended and not creating unintended consequences elsewhere).  

Is it ok to average the left and right measurements in REW and eq it (the average measurement)?

I wouldn't ordinarily recommend it.  Even very well-matched speakers won't be identical, and some kinds of room effects that can be meaningfully eq'ed will be different for each (unless your room is perfectly and totally symmetrical).  A good example of this is corner loading.  If one speaker is in a corner and the other is not, you'll get very, very different bass measurements from the baffle step down to DC.  If you were dealing with a DSP box that had a limited number of filters, I'd say "sure" average the left and right, but when you have essentially infinite EQ, there's really no need to compromise.

I plan to type in the eq filters for my center and main (38 total filters) using jriver parametric eq and type in filters for my side and rear surrounds using parametric eq2 (10 total filters). Is that ok? Is there a max number of filters one can enter in jriver DSP? Can I enter all 48 filters in one parametric eq? Also, will the change take effect immediately by ticking and unticking the parametric eq?  Thanks for all the help as I am new to jriver and REW eq.

Yes, yes, and yes.  
1) There is no limit to the number of filters, and you can put as many as you want in one parametric EQ.  
2) JRiver just offers two PEQs so you put another processing step (like convolution) in between various steps in the audio chain.
3) Ticking and unticking the boxes has an immediate effect, no need to start or stop playback.  Depending on your buffer settings it might take a beat or two, but it engages in real time as soon as it can.

[harlington]

Awesome. Thanks for answering all of my questions. I will eq one channel at a time instead of average two channels. 

[harlington]

One more question: if I add a -6db gain in the filter for all 7.1 channels, does that mean my overall system now is -6db less than what Audyssey sets them to? In other word, my reference level is now with +6 master volume as opposed to 0MV before?  Thanks.

[mwillems]

One more question: if I add a -6db gain in the filter for all 7.1 channels, does that mean my overall system now is -6db less than what Audyssey sets them to? In other word, my reference level is now with +6 master volume as opposed to 0MV before?  Thanks.

Yes you're reducing your overall gain by 6 dB, although you might want to turn off Audyssey until you're done measuring.

[harlington]

^^ are you saying I should turn off Audyssey for measuring and applying eq, put filters in jriver DSP and then turn on Audyssey? How does jriver DSP work? Signal goes in--> jriver processes DSP --> send correction signal to AVR via hdmi? Is that about right? Thanks again.

[mwillems]

Audyysey is doing additional processing that will color the results.  It's another layer of processing in the signal chain.  If you're just using it for volume calibration, it's probably fine to leave it on, but (depending on the system) it can also do significant EQ processing and that isn't necessarily desirable. 

You don't want to wind up "fighting" with audyysey's EQ.

[harlington]

^^ make sense. Man, I just entered >100 filters in jriver. I guess I have to start over   by measuring/ applying eq with Audyssey turned off, then turn it on after all.  I always play movie with Audyssey on. I like what it does to eq my FR and was thinking if I can fine tune it a bit with jriver DSP.  Thanks.
Wait, I am confused now. If I measure and apply eq in jriver with Audyssey off, what happens when I turn it on?  I assume it will try to correct the correction filters in jriver.  Audyssey calibrates speakers independently with receiver tones so it does not see my jriver filter corrections while calibrating. How does jriver DSP work with Audyssey avr?

[mwillems]

I was suggesting turning off audyssey and leaving it off unless you just use it for volume setting.  You're effectively doing in JRiver what Audyssey is doing, but with the potential for much greater precision.  It's hard enough to eq the complex interaction of a speaker and a room without adding a third leg to the tripod and eqing audyssey's interaction with both of those as well.

If you want to keep using Audyssey, then yes absolutely leave it on while measuring (you want to take your measurements in a way that replicates your listening as closely as you can).  I just can't vouch for the results; my advice is to only eq in one place if possible.

[jdubs]

Pretty cool developments with REW to make it more of an all-in-one solution:

http://www.hometheatershack.com/forums/rew-forum/99673-feature-request-frequency-dependent-windowing-8.html

-Jim

[mwillems]

Oh that's pretty neat!  Once it's out of beta, I'll try and update the guide.

[MaximalC]

Thanks, mwillems, for this fantastic tutorial. Separating speaker correction, phase correction, and room correction is a great idea. I do have a question regarding properly aligning impulse responses, though. You've indicated that a measurement's greatest peak should have a positive polarity, however I have a situation where the tweeter and woofer do not agree in this regard. That is to say, the tweeter's greatest peak is on the latter half of the first cycle, while the woofer's greatest peak is on the initial half of the first cycle. I inverted the woofer's impulse response to rectify the upside down peak, however this resulted in only the tweeter having a negative polarity peak preceding the main peak. This also made the phase response of the woofer measurement look like zebra stripes compared to how it did initially.

In such a situation, would it not be preferable to allow the tallest peak of the woofer's impulse response to remain in the negative direction? The peak immediately following the woofer's tallest peak is only slightly shorter, and they're both part of the same wave cycle. I would intuit, based on the improved phase response, that it would be desirable not to invert the woofer's impulse response in this situation, however that would be in violation of your guide, which places some emphasis on ensuring the tallest peak always has a positive polarity.

Another issue I came across with regard to impulse response alignment has to do with the stitching of measurements. The stitching dialogue box includes a checkbox that attempts to automatically align the impulses for maximum accuracy, however using this box results in my woofer impulse being shoved several cycles forward, such that neither of its peaks are anywhere close to aligning with the tweeter peak. The phase response in the transition band of the resulting measurement is also far broader and more irregular than if I had just left the woofer's impulse where it was (with none of the automatic stitching compensation). I only mention this because I think some users may end up with suboptimal results if attempting to always follow your directions inerrantly if they don't fully understanding why they are doing what they're doing.

Then again, perhaps I'm misunderstanding something myself, in which case I gladly welcome your feedback. Regardless, I'd appreciate your confirmation that positive peak polarity is not mandatory and that phase consideration may be best left disabled when stitching if proper impulse alignment has already been established.

[mwillems]

Thanks, mwillems, for this fantastic tutorial. Separating speaker correction, phase correction, and room correction is a great idea. I do have a question regarding properly aligning impulse responses, though. You've indicated that a measurement's greatest peak should have a positive polarity, however I have a situation where the tweeter and woofer do not agree in this regard. That is to say, the tweeter's greatest peak is on the latter half of the first cycle, while the woofer's greatest peak is on the initial half of the first cycle. I inverted the woofer's impulse response to rectify the upside down peak, however this resulted in only the tweeter having a negative polarity peak preceding the main peak. This also made the phase response of the woofer measurement look like zebra stripes compared to how it did initially.

In such a situation, would it not be preferable to allow the tallest peak of the woofer's impulse response to remain in the negative direction? The peak immediately following the woofer's tallest peak is only slightly shorter, and they're both part of the same wave cycle. I would intuit, based on the improved phase response, that it would be desirable not to invert the woofer's impulse response in this situation, however that would be in violation of your guide, which places some emphasis on ensuring the tallest peak always has a positive polarity.

I only suggested ensuring that the peak be positive because some microphones invert, and if all of your impulse measurements wind up upside down it's better for your phase correction to correct the microphone's inversion.  If one of your elements is measuring inverted and the other right-side up and the microphone is facing them from the same angle, your speakers are wired out of phase, and you should not invert the impulse.

Some speakers are deliberately wired out of phase because of the crossover topography used.  For example, a symmetrical Butterworth 2nd order will cause the two elements to be 180 degrees out of phase which would cause terrible cancellation.  The conventional wisdom in that circumstance is to invert one or the other speaker so that they're approximately "in phase."  

Bottom line, if the mic is facing the same direction and one element measures positive and the other negative, correct them as is without inversion, it's probably a design feature of your speaker and you need to take it into account when doing correction.  

I added some clarifying text above.

Another issue I came across with regard to impulse response alignment has to do with the stitching of measurements. The stitching dialogue box includes a checkbox that attempts to automatically align the impulses for maximum accuracy, however using this box results in my woofer impulse being shoved several cycles forward, such that neither of its peaks are anywhere close to aligning with the tweeter peak. The phase response in the transition band of the resulting measurement is also far broader and more irregular than if I had just left the woofer's impulse where it was (with none of the automatic stitching compensation). I only mention this because I think some users may end up with suboptimal results if attempting to always follow your directions inerrantly if they don't fully understanding why they are doing what they're doing.

Then again, perhaps I'm misunderstanding something myself, in which case I gladly welcome your feedback. Regardless, I'd appreciate your confirmation that positive peak polarity is not mandatory and that phase consideration may be best left disabled when stitching if proper impulse alignment has already been established.

I do not recommend using the automatic phase alignment in the stitch dialog, and looking back at the tutorial I don't think I mention it?  I only provided instructions for the other two entries, which is an oversight.  In any event it's not mandatory, and I don't recommend using it.

I do recommend using the separate "match" tool in Holm, but that only level matches, it's doesn't affect phase.  Maybe that's the source of confusion?  I added some clarifying text above.

All that said, if your 1 meter measurement has a drastically different phase picture than your two stitched measurements that may suggest that something is out of whack in the manual alignment.

[MaximalC]

I only suggested ensuring that the peak be positive because some microphones invert, and if all of your impulse measurements wind up upside down it's better for your phase correction to correct the microphone's inversion.  If one of your elements is measuring inverted and the other right-side up and the microphone is facing them from the same angle, your speakers are wired out of phase, and you should not invert the impulse.

Some speakers are deliberately wired out of phase because of the crossover topography used.  For example, a symmetrical Butterworth 2nd order will cause the two elements to be 180 degrees out of phase which would cause terrible cancellation.  The conventional wisdom in that circumstance is to invert one or the other speaker so that they're approximately "in phase."  

Bottom line, if the mic is facing the same direction and one element measures positive and the other negative, correct them as is without inversion, it's probably a design feature of your speaker and you need to take it into account when doing correction.

Wow, thanks for the rapid and thorough response. This is good to know. I need to go back and double-check that I measured and processed properly, because only one of my two speakers had this discrepancy.

I do not recommend using the automatic phase alignment in the stitch dialog, and looking back at the tutorial I don't think I mention it?  I only provided instructions for the other two entries, which is an oversight.  In any event it's not mandatory, and I don't recommend using it.

Ahhh, ok. Not using it concurs with my personal experience as well  .

I do recommend using the separate "match" tool in Holm, but that only level matches, it's doesn't affect phase.  Maybe that's the source of confusion?  I added some clarifying text above.

I'm glad you mentioned the "match" tool, as that has been my second source of confusion today. I understand that the match tool is intended to compensate for level discrepancies between measurements, and that by compensating the gain of one measurement, it should transition more seamlessly with the other. This worked fine and dandy for my 1 ft measurements, which required very meager adjustments and produced a fairly flat stitched response.

However, with the 3 inch measurements, the match tool suggested a 2.5 dB boost of the woofer response (when attempting to match at the crossover region as suggested). This resulted in a substantial bass boost in the stitched response, and I'm skeptical that it is an accurate representation of the holistic speaker response at 3 inches. I have a screenshot below so you can see what I mean.


The blue trace is the stitched response at 1 foot and the green line is the stitched response at 3 inches (the red trace is a measurement at tweeter height 2 ft away and can be ignored for the purpose of this discussion). I think the problem may be that the directional higher frequencies from the tweeter are not diffracting to the woofer measurement as well as the lower frequencies are dispersing into the tweeter measurement. When Holm attempts to match level around my speakers' 1725 Hz crossover frequency, it interprets the woofer measurement to be underpowered (due to the lower measured volume above the crossover) and overcompensates to solve a problem that may not exist. The stitched response would be much more flat if I didn't match levels before stitching.

Unfortunately, Holm is automatically matching level whenever I use the stitch command; I don't even have to match level manually prior. Seeing as how there is no way to disengage level matching during stitching, I tried a wonky workaround by transitioning at an alternative frequency band, but I don't know if that's really proper considering it's not the true crossover frequency.

My worry is that if I average the responses as they are and attempt to EQ my speaker, everything below 1 kHz is going to be scooped out unnecessarily. I'm almost tempted to go exclusively based on the 1 foot response, but there are already some room effects creeping into the low end at that spot, so that's not an ideal solution either. I'm wondering if backing out to 4 or 5 inches instead of 3 will provide a more even response for the woofer measurement and negate the exaggerated level matching. Am I even correct in thinking the 3 inch response (green trace) is objectionable in the first place?

[mwillems]

What settings did you use for the "match" tool to generate your green line (frequency and width)? If I had to guess, I would say that you used too narrow of a width and you'd get a more even match if you broadened it.  But let me know what parameters you used.

Your three inch measurements should not be that different from your 12 inch; a slight rise in the midbass region is expected because a variety of room cancellations will be missing, but it shouldn't be that much and that that flat.  You can try measuring farther out, but you can probably just tweak the level matching width to get the two more closely matched.  That goes for the stitch as well; you can manipulate how it behaves by adjusting the width up or down.

I've only ever seen the kind of thing you're seeing when the width was narrow and the specified region had a trend in it that was not continued outside of the affected region (effectively fooling the algorithm).

[MaximalC]

What settings did you use for the "match" tool to generate your green line (frequency and width)? If I had to guess, I would say that you used too narrow of a width and you'd get a more even match if you broadened it.  But let me know what parameters you used.

Your three inch measurements should not be that different from your 12 inch; a slight rise in the midbass region is expected because a variety of room cancellations will be missing, but it shouldn't be that much and that that flat.  You can try measuring farther out, but you can probably just tweak the level matching width to get the two more closely matched.  That goes for the stitch as well; you can manipulate how it behaves by adjusting the width up or down.

I've only ever seen the kind of thing you're seeing when the width was narrow and the specified region had a trend in it that was not continued outside of the affected region (effectively fooling the algorithm).

I used a 1380 Hz width for my speakers' 1725 Hz crossover frequency (was aiming for 75-80% of the crossover frequency as alluded to in your tutorial). I have since tried using a broader width (equal to the crossover frequency), and while this did reduce the boost a little bit, it also resulted in a broader scoop in the transition band. I still felt that the woofer boost was exaggerated quite a bit, so I have since abandoned attempting to use the 3 inch measurements.

Instead, I painstakingly measured every 1-inch increment between 3 and 12 inches, imported them into REW and matched their overall levels manually to more closely compare the tradeoffs between broadband response and room effects. Ultimately, I found that measurements between 8 and 10 inches yielded the best compromise between balanced bass/treble and minimal room effects. My plan now is to average those three measurements and use that for my speaker EQ baseline.

Speaking of EQ, would it not be preferable to use REW's "variable" smoothing algorithm as opposed to a strict 1/12 octave resolution? I think I read a post from John Mulcahy a while back saying that it would even be justified to disable smoothing for the purpose of bass correction, whereas the uppermost frequencies could do fine with only 1/3 octave resolution. Is there any advantage to using a consistent smoothing resolution such as 1/12 octave?

[mwillems]

Speaking of EQ, would it not be preferable to use REW's "variable" smoothing algorithm as opposed to a strict 1/12 octave resolution? I think I read a post from John Mulcahy a while back saying that it would even be justified to disable smoothing for the purpose of bass correction, whereas the uppermost frequencies could do fine with only 1/3 octave resolution. Is there any advantage to using a consistent smoothing resolution such as 1/12 octave?

I think you might be misunderstanding how frequency dependent windowing (FDW) works.  It's not the same as conventional 1/12 octave smoothing, which you can see if you switch the smoothing algorithm in Holm to use the regular smoothing instead.  The point of FDW is to have a variable gate that changes with frequency to gate out late reflections without rendering any part of the graph unreliable (which is what happens with a fixed gate).  The effect of FDW (depending on the width of the gate) looks like smoothing, but the smoothing is caused by gating out late reflections not by mathematical averaging.  So the effects of FDW are in most cases a "truer" reflection of the signal than smoothing would be.  It's not comparable in function to conventionall 1/12 octave smoothing.

REW's variable smoothing is an imperfect attempt to solve the same problem that FDW solves (i.e. scaled variable smoothing), so in my view it's not preferable. REW just recently implemented FDW, so I suspect there was a felt need for it over there too.

Reasonable minds can differ about how wide the FDW "floating gate" should be (12, or 6, or 9, or 15). You can see for yourself where the "gate" is for a few sample frequencies on the impulse response graph if you want to get a sense for what effect the parameter has.  But the whole point of FDW is that you don't have to use a different approach for different frequencies because it scales with frequency. The latest betas of REW now support FDW, but it was not supported when this guide was written and I haven't tested the functionality in REW.

[MaximalC]

I think you might be misunderstanding how frequency dependent windowing (FDW) works.  It's not the same as conventional 1/12 octave smoothing, which you can see if you switch the smoothing algorithm in Holm to use the regular smoothing instead.  The point of FDW is to have a variable gate that changes with frequency to gate out late reflections without rendering any part of the graph unreliable (which is what happens with a fixed gate).  The effect of FDW (depending on the width of the gate) looks like smoothing, but the smoothing is caused by gating out late reflections not by mathematical averaging.  So the effects of FDW are in most cases a "truer" reflection of the signal than smoothing would be.  It's not comparable in function to conventionall 1/12 octave smoothing.

REW's variable smoothing is an imperfect attempt to solve the same problem that FDW solves (i.e. scaled variable smoothing), so in my view it's not preferable. REW just recently implemented FDW, so I suspect there was a felt need for it over there too.

I apologize for conflating FDW and traditional smoothing above. I was fixated on octave resolution without discriminating the methodology, even though I understand and agree that FDW is preferable. I suppose a better question would be, are there any advantages to using both methods in tandem? For example, if I used 1/48 octave FDW and then variable smoothing in REW, would that not offer the best of both worlds? My understanding is that the advantage of variable smoothing is not simply to gloss over comb filtering, but to portray the frequency response in a way that more closely approximates our ability to discriminate tones in a given frequency range. If our ears can't discriminate beyond 1/3 octave resolution above 10 kHz, why preserve a higher resolution trace of those frequencies for the purposes of EQ? I guess I'm trying to err on the side of caution by not invoking filters to solve problems that we can't hear, even though it's true that they exist.

It would be nifty if we could enjoy high resolution low frequencies with minimal room effects (courtesy of FDW) without over-correcting for high frequency variations that we can't hear (courtesy of variable smoothing). I hadn't been regarding the two processes as mutually exclusive because their purposes, while related, aren't quite the same. I guess the ultimate arbiter would be to try it and see how it sounds, but I always prefer to have a sound theoretical basis before deferring to my own subjective impression.

Thanks again for your rapid feedback. If only professional tech support were as timely! 

[mwillems]

I've heard the view expressed that we can't hear variations finer than 1/3 octave at high frequencies, but my experience has been that it's a matter of degree.  High Q, high intensity peaks or dips can be quite audible even when very narrow in my experience.  

For example, I happen to have a driver with an 8dB resonant peak at 17.5KHz that's only about 1/6 octave wide.  If I don't correct that peak, I can hear it even though my hearing is very poor that high up.  In part I notice because it's resonant and distorts like nobody's business when uncorrected, but I also notice it because it causes lots of intermod farther down the frequency spectrum.  It's clearly audible when uncorrected, not at all subtle.

So I'm of the view that there are potential advantages in fixing narrow band high frequency peaks and dips, especially of the high Q, high intensity variety.  I think there's an argument for skipping high frequency correction with conventional DSP systems (like a DCX2496 or a miniDSP) that have limited processing power, because you need to prioritize bang for the buck.  But given that JRiver's PEQ bank is effectively infinite and not particularly CPU intensive, there's certainly no harm in correcting all peaks and dips provided they're really a feature of your driver's response and not of the room.  

I think an empirical approach is the answer; with JRiver's DSP you can flick filters on and off at will and see if you can hear it.  If you want a scientific approach, have a third party toggle it out of sight, DSP changes take effect seamlessly in real time during playback, so it's not hard to do a single-blind A/B test.

[mattkhan]

It would be nifty if we could enjoy high resolution low frequencies with minimal room effects (courtesy of FDW) without over-correcting for high frequency variations that we can't hear (courtesy of variable smoothing).

You can. Nothing says the no of cycles to use for the fdw size has to stay constant across the pass band.

I am not aware of any freeware that implements this though, it is a feature found in acourate which does actually combine fdw with a proprietary smoothing algorithm when preparing a response for correction. Ultimately you will just need to try some variations out and see what you prefer.

What is the design of your speakers BTW? You may find that the nature of the speaker influences the optimal correction. In particular I am thinking of the nature of off axis performance and whether you have multiple driver arrays.

[MaximalC]

You can. Nothing says the no of cycles to use for the fdw size has to stay constant across the pass band.

I am not aware of any freeware that implements this though, it is a feature found in acourate which does actually combine fdw with a proprietary smoothing algorithm when preparing a response for correction. Ultimately you will just need to try some variations out and see what you prefer.

What is the design of your speakers BTW? You may find that the nature of the speaker influences the optimal correction. In particular I am thinking of the nature of off axis performance and whether you have multiple driver arrays.

That's interesting, I'll have to look into Accourate. My speakers are the JBL LSR305, which have a 5 inch woofer mounted below a 1 inch tweeter. The tweeter's waveguide is supposed to offer exceptional off-axis response.

[MaximalC]

Alright, I've completed the direct sound EQ portion of my speaker correction journey and have moved on to rePhase. Because nothing is ever as straightforward as it seems on paper, I've encountered another bugaboo. My stitched speaker's phase response wraps twice, once in the low end, and once in the midrange. The low end wrap was resolved by inserting the proper port frequency and type into rePhase; great! Since I know my speaker uses a 1725 Hz 4th order Linkwitz-Riley crossover filter, I figured I'd remedy the remaining phase wrap by picking the 48 dB/octave L-R linearization filter. Contrary to expectation, this merely flipped the phase wrap (so that the slope of the phase was now leaning to the right instead of the left). Also contrary to expectation, replacing the 4th order LR filter with a 1st order LR filter completely eliminated the remaining phase wrap... How is this possible considering I know this speaker employs a 4th order LR filter (I even double-checked the manual to make sure I wasn't mistaken)?

Is the stitching bandwidth relevant to what I'm observing here? Should I be calculating the stitching window based on my speaker's crossover slope instead of approximating one octave? Or is that completely irrelevant?

Should I just be content that my phase has flattened and move on, or should I not trust what I'm seeing in rePhase considering I know the chosen filter does not correspond to my speaker? Also, does it matter whether I do phase correction before EQing instead of after? I assumed the difference would be negligible considering I used very minor EQ adjustments, but perhaps I assumed wrong...

Sorry for so many questions, I'm a total phase n00b  .

[mwillems]

Alright, I've completed the direct sound EQ portion of my speaker correction journey and have moved on to rePhase. Because nothing is ever as straightforward as it seems on paper, I've encountered another bugaboo. My stitched speaker's phase response wraps twice, once in the low end, and once in the midrange. The low end wrap was resolved by inserting the proper port frequency and type into rePhase; great! Since I know my speaker uses a 1725 Hz 4th order Linkwitz-Riley crossover filter, I figured I'd remedy the remaining phase wrap by picking the 48 dB/octave L-R linearization filter. Contrary to expectation, this merely flipped the phase wrap (so that the slope of the phase was now leaning to the right instead of the left). Also contrary to expectation, replacing the 4th order LR filter with a 1st order LR filter completely eliminated the remaining phase wrap... How is this possible considering I know this speaker employs a 4th order LR filter (I even double-checked the manual to make sure I wasn't mistaken)?

Is the stitching bandwidth relevant to what I'm observing here? Should I be calculating the stitching window based on my speaker's crossover slope instead of approximating one octave? Or is that completely irrelevant?

Three points:

1) A fourth order Linkwitz-Riley is a 24dB filter, not a 48dB filter (which is eighth order), so that's a significant portion of your issue.  Using 48 instead of 24 will introduce 360 degrees of unnecessary phase rotation.  What you describe (the wrap just changing directions) sounds exactly like what I would expect from that.  A 4th order LR introduces 360 degrees of rotation, and eight order introduces 720 degrees.  So by correcting for an 8th order, you're correcting the 360 from the 4th order, and then rotating another bonus 360 degrees.  In essence you're going from -360 out of phase to +360 out of phase (putting you back where you started but in the opposite direction).
2) There normally isn't such a thing as a "first order" Linkwitz-Riley, so I'm not sure how to advise you about that piece. Can you describe that filter a bit more?  What's the filter slope?
3) Unless the crossover is applied to speaker elements that are perfectly flat in terms of frequency and phase throughout the crossover transition band, just undoing the phase wrap from the crossover won't solve 100% of the phase wrap, because the elements themselves will introduce some phase wrap from their own roll off.  So you may need to do some additional tweaking even after you dial in the correct crossover, or you may need less correction depending on the elements.  It may be that the designer already factored that in and just created an effective 4th order LR slope taking the drivers' roll-off into account. The spec sheet for your speakers refers to the crossover as a 4th order acoustic crossover, which may mean that there's an effective 4th order LR slope taking the drivers into account, which would make your life easier if that's what that means.

Should I just be content that my phase has flattened and move on, or should I not trust what I'm seeing in rePhase considering I know the chosen filter does not correspond to my speaker?

Try my advice above and see what happens.  Ultimately, if the filter is flat, and the speaker is flat when you measure it after applying the filter, then mission accomplished. Declare victory and move on  

Also, does it matter whether I do phase correction before EQing instead of after? I assumed the difference would be negligible considering I used very minor EQ adjustments, but perhaps I assumed wrong...

If you used the raw measurements as the basis for your phase correction, I would advise putting the convolution module first (before EQ).  If you used your corrected measurements for your phase correction I would advise placing the convolution module after the EQ. Convolution is one of the only modules where the order can actually matter.  I'm not sure that it will make a huge difference, but I would advise putting the filter in the same place in the "chain" that you developed it if that makes sense.

[MaximalC]

1) A fourth order Linkwitz-Riley is a 24dB filter, not a 48dB filter (which is eighth order), so that's a significant portion of your issue.  Using 48 instead of 24 will introduce 360 degrees of unnecessary phase rotation.  What you describe (the wrap just changing directions) sounds exactly like what I would expect from that.  A 4th order LR introduces 360 degrees of rotation, and eight order introduces 720 degrees.  So by correcting for an 8th order, you're correcting the 360 from the 4th order, and then rotating another bonus 360 degrees.  In essence you're going from -360 out of phase to +360 out of phase (putting you back where you started but in the opposite direction).

Well don't I feel silly...  I had been thinking that 1st order was 12 dB/octave.  That certainly does explain a good deal of my confusion!

2) There normally isn't such a thing as a "first order" Linkwitz-Riley, so I'm not sure how to advise you about that piece. Can you describe that filter a bit more?  What's the filter slope?

It was 12 dB/octave (2nd order). My bad!

3) Unless the crossover is applied to speaker elements that are perfectly flat in terms of frequency and phase throughout the crossover transition band, just undoing the phase wrap from the crossover won't solve 100% of the phase wrap, because the elements themselves will introduce some phase wrap from their own roll off.  So you may need to do some additional tweaking even after you dial in the correct crossover, or you may need less correction depending on the elements.  It may be that the designer already factored that in and just created an effective 4th order LR slope taking the drivers' roll-off into account. The spec sheet for your speakers refers to the crossover as a 4th order acoustic crossover, which may mean that there's an effective 4th order LR slope taking the drivers into account, which would make your life easier if that's what that means.

Well I'll be... The "acoustic crossover" distinction appears to be an important one. Per Wikipedia: "A third- or fourth-order acoustic crossover often has just a second order electrical filter. This requires that speaker drivers be well behaved a considerable way from the nominal crossover frequency, and further that the high frequency driver be able to survive a considerable input in a frequency range below its crossover point." This would appear to be precisely my situation. Since the speaker drivers are already contributing to the stated 4th order crossover slope acoustically, that would explain why I only needed a 12 dB/octave phase correction in rePhase to account for the 2nd order electrical filter.

Thank you!

[mattkhan]

Oh that's pretty neat!  Once it's out of beta, I'll try and update the guide.

REW 5.13 is hot off the press and has an FDW option as well as an equivalent rectangular bandwidth smoothing option (download from http://www.hometheatershack.com/forums/downloads-area/19-downloads-page.html)

[mwillems]

REW 5.13 is hot off the press and has an FDW option as well as an equivalent rectangular bandwidth smoothing option (download from http://www.hometheatershack.com/forums/downloads-area/19-downloads-page.html)

Aces.  It'll probably take me a bit to integrate it (I'll probably just note it as an alternative rather than doing a wholesale revision, but I'd like to fiddle with it and see how it does).

[PrinterPrinter]

Hello,
Thanks for the guide!

I'm about to start with his - trying to add room correction to a 2.1 system without dedicated sub channel - so I'm EQing the whole system together.
I wonder if the UMIK is still the recommended Calibrated Mic for this kind of thing? Any other suggestions are welcome, thanks!

[mwillems]

I wonder if the UMIK is still the recommended Calibrated Mic for this kind of thing? Any other suggestions are welcome, thanks!

Any properly calibrated mic should work fine for basic correction of the type outlined in this guide.  EQing a sub as part of the whole system will be a little tricky as the biggest problems are often delay related and you won'be able to correct that withot a dedicated channel.

[5150]

I've been working on a guide to active speakers/crossover design in JRiver, but I have an infant daughter so the guide has languished unfinished on my "to do" pile. 

I'm a little late to the part, but I'm really interested in the topic and anxiously await your active guide!  Thanks for the great contributions you've already made!

[jomal]

Hello everybody.

Is there a freeware frequency- and impulse-response measurement tool like HOLMImpulse for Mac / OS X?

Thanks in advance.