convolution room correction and room curve

[linuxonly]

@ddude003

Following this tutorial

I expected R corrected to bear a more downward slope. Am I missing something?



I did not use the Harman curve but a log progression instead

 

[moedra]

@ddude003

Following this tutorial

I expected R corrected to bear a more downward slope. Am I missing something?

View attachment 72926

I did not use the Harman curve but a log progression instead

View attachment 72927

Something doesn't look right. Your inverted response should all be below zero. It won't necessarily look like it's right, but if you multiply the inversion by the original response, you should get the corrected simulation. Put your mdat file up here and I'll look at what you have when I get a minute.

 

[linuxonly]

Something doesn't look right. Your inverted response should all be below zero. It won't necessarily look like it's right, but if you multiply the inversion by the original response, you should get the corrected simulation. Put your mdat file up here and I'll look at what you have when I get a minute.

Exactly my thought.

I did a second tentative. Better results as I forgot some steps in the first but still the same problem.


Thanks a lot @moedra

 

[ddude003]

Sorry, without seeing each step along the way it is not clear exactly what is going on... Don't get to see the inversions... What happens if you use a house curve? Same issues?

 

[linuxonly]

What happens if you use a house curve? Same issues?

Hi,

Thanks for having a peek.

Yes, exactly the same issue with the Harman curve. Sole difference, it has more bass. Maybe a bug or I forgot a crucial step. The tutorial is made with an earlier version of REW and there are differences.
As soon as I click on `Calculate target from response`, the target curve moves down and most of the corrction is therefore above it.

This does not happen when I generate EQ filters. The predicted output follows the home curve.



However, while the correction is being made, this not does sound right using EQ. The stereo image of a singer is totally lost. I.e. the voice seems to come from the speakers instead of a point between them. My bad, I did not study how to properly EQ. Only testing the EQ panel here do determine why the room curve is applied to filters and not to impulse file.

 

[moedra]

I honestly have no idea how the inversions/calculations resulted in what you have there. Haven't been through that video. I do have my own method, which I can apply to this and we'll see what we get.

 

[John Mulcahy]

As soon as I click on `Calculate target from response`, the target curve moves down and most of the corrction is therefore above it.

Calculate target level is something that is done before making corrections, not afterwards. It asks REW to compare the response to the target and pick a target level that puts the target roughly in the middle of the response.

 

[linuxonly]

Calculate target level is something that is done before making corrections, not afterwards. It asks REW to compare the response to the target and pick a target level that puts the target roughly in the middle of the response.

That's what I did. And I don't see the expected effect whatsoever. Clearly something has changed in REW between the version the tutorial was made and the current version. Whenever I follow the tutorial to the letter for the correction, it happens. The author method for the calculation is to do arithmetic calculation. Here's the specific portion of his video:

L / Target LR = A over B (L is curve 2, Target LR is curve 4))
1 / A = L inverted
apply minimal phase to L inverted (L inverted MP) -> this is the impulse filter to export (curve 8)
proof made by multiplying L inverted MP * L to get what's expected, L corrected that is (curve 9). And it does not work for me after many tentatives on 2 systems.

Repeat for R

I don't know why he proceeds that way, in 2 steps.

This seems to produce proper data using single step (my test):
Target LR / L = L inverted (limit to 20 -20kHz or whatever, regularisation 8% or 5dB)
apply minimal phase to L inverted (L inverted MP) -> this is the impulse filter to export
proof made by multiplying L inverted MP * L to get what's expected, L corrected that is (curve 11). And this one does follow the room curve. Might have unwanted boosts, not tested yet, only theory here.

Repeat for R

Makes sense?

 

[moedra]

Try this out. I know it basically ignores and sidesteps your questions, but I'm curious to know what results it gives you. There are two filter sets there: measurements 6/7, and measurements 10/11. You can export each of them to check how they sound. In my tests I have noticed that you sometimes have to check the "export minphase version" in the export window to avoid pre-ringing. This seems to be an issue when exporting the "measured" version even if I have previously made the inverted responses minimum phase copies. For some reason checking the minphase version box upon export fixes that. I don't quite understand it.

 

[linuxonly]

Try this out. I know it basically ignores and sidesteps your questions, but I'm curious to know what results it gives you. There are two filter sets there: measurements 6/7, and measurements 10/11. You can export each of them to check how they sound. In my tests I have noticed that you sometimes have to check the "export minphase version" in the export window to avoid pre-ringing. This seems to be an issue when exporting the "measured" version even if I have previously made the inverted responses minimum phase copies. For some reason checking the minphase version box upon export fixes that. I don't quite understand it.

Thanks. I appreciate your effort. Will happily check that later today (currently 3:37 AM)


Care to explain the help file?


As you noted, I sometimes tick a different option to export as wav. It's because I don't understand the interface and the help does not explain why it is like that. I don't understand when I export L inverted MP and R inverted MP for example, why I am being asked to select whether to export a measured or MP version of the IR. Should it not be obvious? If I wouldn't, I'd had exported L inverted and R inverted, wouldn't I?

Export the impulse response for the selected measurement(s) in WAV format, written as mono or stereo data. The sample format can be chosen as 16, 24 or 32-bit signed PCM or 32-bit Float. Float is recommended if the application using the data can accept this, particularly if the response is not normalised.

The dialog provides options to choose whether to export the measured IR, the IR after any EQ filters have been applied to it or a minimum phase version of the IR.

What would be the advantages and inconvinients of normalizing?

And why place t=0 at a specific index? I use timing references when I record the sweeps. Shouldn't it be automatic since it's clearly identifiable? And which specific index would it be? I need that to be time and phase aligned with the sound, that's all.

The response can optionally be normalised so that the peak value is unity (0 dBFS). There is also an option to apply the current impulse response window settings to the response before exporting it and an option to place t=0 at a specific sample index in the export. The export can be limited to a specific sample count if desired.
.

Best regards.

 

[linuxonly]

There are two filter sets there: measurements 6/7, and measurements 10/11.

Oh thats' interestiing. I exported measurements 6-7 and 10-11 and there are currently under evaluation.

I noticed your special target file and also noticed you made your arithmetic calculations with Target MP / L and Target MP / R so generated one (curve 21).
Also noticed the smoother downward slope than mine and the fact everything inversed is under 0 dB.

Attaching your results with my last calculations from yesterday appended for your comparison (traces 13-20).

After I'm done with the home theater testing (in stereo mode), will try to apply your method on my other computer with analog stereo connected to full range speakers. Anything in particular I should be aware of?

Thans again.

 

[moedra]

I don't let anything get boosted in the filter. It is safer to allow the filter to attenuate everything, and compensate for the loss of volume post-filter so that it can be controlled. This is what I was taught by Bob Katz a few years ago. His instruction was to place the target underneath the smoothed response (touching the dips) and run the calculations from there. I also keep an eye on large dips below the room's transition zone and ignore them if they are too deep. I don't think you had that problem.

Some things to keep in mind, sure:

There is no point in generating a single average response of the left and right speakers. Ideally we are correcting them individually, so I want to treat them on their own and generate filters independently rather than from an average of the left and right.

I smoothed everything to 1/12 smoothing and created a minimum phase version of the target and both responses before I inverted anything. It looks like I deleted the notes for the target, which I tend to do because I like to type my own, only in this case I seem to have forgotten to type anything. But if you look at the target's phase you'll see that it does indeed have a minimum phase response.

Bob told me not to use the FDW because REW doesn't exactly do it right for the purpose of creating correction filters. It took me a long time to understand what he meant by that, but having studied Acourate for a couple of years and analyzing how macro 1 works, I now think I get it. From my understanding of it, Acourate smooths the response in two steps inside of the first macro. First it applies a psychoacoustic treatment that eliminates certain dips, and then it applies the FDW afterward. The psychoacoustic treatment helps the FDW avoid calculating those dips and creating unnatural-looking downward spikes. As far as I can tell, there is no substitute for Acourate's psychoacoustic treatment in REW. Maybe there is a way to simply hack off those dips prior to applying the windowing, but that would probably not be psychoacoustically optimal and it is much easier to smooth everything, including the target, to 1/6 or 1/12 smoothing before you calculate anything.

What is it that you did in the new file? I see the new calculations under mine, but I don't know what you did. Have you got a procedure for that?

 

[linuxonly]

What is it that you did in the new file? I see the new calculations under mine, but I don't know what you did. Have you got a procedure for that?

Hi,

Thanks for the info. Much appreciated.

It is not something new I attached. When I realised something could be way off with my august 14 calculations, I tried another method on august 17. What you see is the result of that. It's only there to show it's very near to what you've sent me the 18th. Instead of using the 2 step method for calculating the inverted responses as I did the 14th, I used a 1 step method, like you did but still with a 5dB limitation and a 20-20k response. The corrected response is closer than your results than they were the 14th. E.g. compare your 'L Simulation' with my 'L corrected':



All in all, all those corrections, once imported in EA Convolver sound almost identical. The difference is not very noticeable. Maybe a little more bass, but all resonances are gone. Used several audio test files I know well, I always use the same so I know what to expect and what to watch.

 

[moedra]

Seems I am actually getting better results by not converting the original responses to minimum phase before inversion like I was doing. The target, on the other hand, does need to be converted. Invert the raw responses over the minimum phase target. Leave the inversions as they are. Don't convert them. Upon export, tick the minphase version button as before, rather than measured.

To calculate the simulation, try using the exported filter (import it back into REW) instead of the inverted responses used to make the filter.

 

[linuxonly]

Hi,

Thanks for the details.

I tried to follow your instructions. I'm joining a screen capture in case you got something totally different with the attached mdat file.

- I think the EQ settings for the target curve are important, therefore I'm joining a screen capture
- I had to offset the re-imported impulse files by -114dB. Is that expected?

It's not really important to me the HF cutoff frequency, I'm deaf to frequencies above 8kHz due to age. Guests might find the sound awkward though.

 

[moedra]

Yes I had to lower mine as well. -117dB got me there, but I was using my own file so I could listen to results. It looks like you used the raw target? If so, you should create a minimum phase version of the target and use that instead. The raw target has no phase data, and I think that the end result of using a minimum phase version of the target sounds better.

 

[linuxonly]

Interesting. Thanks. I think I'm already using a MP target. In fact it's a copy of your file I think unaltered.

Weird phase shift in the bass area but I was told to make measurements at 48kHz, from DC to 24kHz.

That's why I posted a screenshot of my EQ settings, in case something was off. I'm not qualified enough to spot that.



1/12 smoothing is musical note by note, right?

Out if curiosity i googled Bob Katz. That man is a legend.

 

[moedra]

I think that the phase shift is low enough that it doesn't matter. The sound of the filters generated by this is really good. The phase of the original response is inverted as well as the magnitude.

 

[linuxonly]

The sound of the filters generated by this is really good

I would agree.

I applied this method on my stereo as well (analog 2.0) and I heard things I have never heard since I moved here but having a good acoustic memory, remembered that they should be there. And the original response was all over the place. So the first step was to rebalance the 3 stereo amplifiers driven by the electronic crossover.

Now you completely forget you're listening to a stereo system, you hear voices, instruments, effects. Amazing!

 

[moedra]

I've got a new procedure for you. I just worked it out this morning and it's sounding great!

This is different than everything else I've seen out there, so bear with me. The goal here for me was to figure out a way to emulate macro 1 and macro 3 in Acourate as much as possible with REW. Macro 1 applies a psychoacoustic treatment to the response before the frequency-dependent window. Essentially, it removes the notch-like dips in the response, smoothing out the underside, before applying the FDW. If you apply the windowing in REW without any sort of treatment like that, the resulting response is not going to be true to its smoothed response, which is more accurate where filters are concerned. This is what Bob was trying to tell me. It is because in Room EQ Wizard, FDW will account for the dip "notches" that we need it to ignore since it doesn't apply the calculations to anything other than the original unsmoothed data... at least from what I can tell. Windowing in REW is always calculated based on the unsmoothed data containing the notches. Furthermore, the inversion is also calculated based on the unsmoothed data. The filters we export from these methods also appear to be unsmoothed, which isn't psychoacoustically treating anything. Acourate, on the other hand, does its inversion to the psychoacoustically smoothed data in macro 3.

I wanted the calculations in REW to be applied to my smoothed data rather than the unsmoothed data in the same way.

I solved this problem by first exporting the original responses as text files. In the export measurement as text dialog I applied the 1/12 smoothing. Applying it here bakes it into the measurement.

Afterward I imported the text responses back into REW, which came in at 1/48 smoothing. Not a big deal, since they are baked/smoothed at 1/12 anyway. Next, I made minimum phase copies of them and deleted the imports. That left me with fundamentally unsmoothed data that was identical to the original responses with 1/12 smoothing applied. Now the inversion calculations would be applied to the 1/12 data rather than the unsmoothed data, which would then result in unsmoothed 1/12 filters. That was the goal, after all. So I divided my minimum phase target (measurement 3) by my new minimum phase copies of the imports (measurements 4 and 5), and exported the resulting filter set (measurements 6 and 7) as a stereo IR wav as usual, ticking the export minphase version button upon export. This filter contains none of the ringing/resonant anomalies I was hearing before, and is very close to the sound of the Acourate filter. The other interesting thing here is that I didn't have to worry about FDW, since the 1/12 smoothing is excellent as-is.

Note the placement of the target and how I have it touching the responses around 70Hz, 100Hz, and 250Hz. It's important to get it sitting right underneath the response like this. This will always be case-specific. However, if you have a large interference null or two below the transition zone, you need to ignore them. In those cases, place the target at the level of the next-lowest dips and allow those interference nulls to drop below the target.

My target is sloped at 0.4 dB per octave for both high and low slope settings.

Also, note that I do not let anything invert over zero. No boosts (i.e. max gain equals 0.0).

 

[linuxonly]

Amazing. Generated one instance for my home theatre and evaluating...

Note the placement of the target and how I have it touching the responses around 70Hz, 100Hz, and 250Hz. It's important to get it sitting right underneath the response like this.

I'm right above the target curve as well, no boost. This home theater is pretty linear to begin with.

My stereo is more problematic. I have a huge drop between 2 kHz and 4 kHz on one channel. I'm not sure what to do with that. Right now, I'm just ignoring it.

 

[moedra]

Amazing. Generated one instance for my home theatre and evaluating...


I'm right above the target curve as well, no boost. This home theater is pretty linear to begin with.

My stereo is more problematic. I have a huge drop between 2 kHz and 4 kHz on one channel. I'm not sure what to do with that. Right now, I'm just ignoring it.

View attachment 73068

You're still boosting during the inversion, which is giving you approximately 7 or 8 decibels of boost above 0dB around 2600Hz. The region between 2000Hz and 4500Hz is boosted. There's a little boosting above 0dB at 1kHz as well. When you do the inversion, leave the boost amount (max gain) at 0.0 like I have in the picture. This is somewhat vital to avoid potential problems clipping the output of the convolver running the filter. Look at mine again. There's nothing going over zero.

 

[linuxonly]

Gotcha, thanks!

IMHO, this method should be pinned.

 

[moedra]

Gotcha, thanks!

IMHO, this method should be pinned.

View attachment 73072

How do your filters sound now?

 

[JStewart]

I've got a new procedure for you. I just worked it out this morning and it's sounding great!

This is different than everything else I've seen out there, so bear with me. The goal here for me was to figure out a way to emulate macro 1 and macro 3 in Acourate as much as possible with REW. Macro 1 applies a psychoacoustic treatment to the response before the frequency-dependent window. Essentially, it removes the notch-like dips in the response, smoothing out the underside, before applying the FDW. If you apply the windowing in REW without any sort of treatment like that, the resulting response is not going to be true to its smoothed response, which is more accurate where filters are concerned. This is what Bob was trying to tell me. It is because in Room EQ Wizard, FDW will account for the dip "notches" that we need it to ignore since it doesn't apply the calculations to anything other than the original unsmoothed data... at least from what I can tell. Windowing in REW is always calculated based on the unsmoothed data containing the notches. Furthermore, the inversion is also calculated based on the unsmoothed data. The filters we export from these methods also appear to be unsmoothed, which isn't psychoacoustically treating anything. Acourate, on the other hand, does its inversion to the psychoacoustically smoothed data in macro 3.

I wanted the calculations in REW to be applied to my smoothed data rather than the unsmoothed data in the same way.

I solved this problem by first exporting the original responses as text files. In the export measurement as text dialog I applied the 1/12 smoothing. Applying it here bakes it into the measurement.

View attachment 73064

Afterward I imported the text responses back into REW, which came in at 1/48 smoothing. Not a big deal, since they are baked/smoothed at 1/12 anyway. Next, I made minimum phase copies of them and deleted the imports. That was necessary because the imported responses don't come in with any phase data and we can't use them for anything. I wanted minimum phase copies anyway, so it worked out. That left me with fundamentally unsmoothed data that was identical to the original responses with 1/12 smoothing applied. Now the inversion calculations would be applied to the 1/12 data rather than the unsmoothed data, which would then result in unsmoothed 1/12 filters. That was the goal, after all. So I divided my minimum phase target (measurement 3) by my new minimum phase copies of the imports (measurements 4 and 5), and exported the resulting filter set (measurements 6 and 7) as a stereo IR wav as usual, ticking the export minphase version button upon export. This filter contains none of the ringing/resonant anomalies I was hearing before, and is very close to the sound of the Acourate filter. The other interesting thing here is that I didn't have to worry about FDW, since the 1/12 smoothing is excellent as-is.

View attachment 73067

View attachment 73065

View attachment 73066

Note the placement of the target and how I have it touching the responses around 70Hz, 100Hz, and 250Hz. It's important to get it sitting right underneath the response like this. This will always be case-specific. However, if you have a large interference null or two below the transition zone, you need to ignore them. In those cases, place the target at the level of the next-lowest dips and allow those interference nulls to drop below the target.

My target is sloped at 0.4 dB per octave for both high and low slope settings.

Also, note that I do not let anything invert over zero. No boosts (i.e. max gain equals 0.0).

Thank you for this! I’ll give it a go first chance I get.

Questions...

I see your measurements start at 0Hz. Any concerns or advice on potential woofer damage? @linuxonly measurements start at 20Hz. Is this also acceptable?

Would you see any value to applying different smoothing to different parts of the frequency range. 1/3 octave smoothing above the “transition“ zone, for example.
Using your method and a bit of copy and paste to exported files, I think this could be accomplished.