Excess group delay and mismatched phase response

[Davide]

Following all comments here, I tried 10 different filters in DLBC to see if one would stand out from the crowd. It is hard with my knowledge to decide. GD is true only for one mic position meaurement? And for one channel at a time? DLBC filter is calculated on an average of 13 measurements. Why to prioritize one GD measure against 13?
How to choose between this filter results for the Front left channel

View attachment 409715

and this other filter?
View attachment 409716

I can tell you that I also get significantly different results by remaking the measurement and filter creation process several times.
Yet even if the microphone positions are not repeatable to the millimeter, the final filter should not differ so much, IMO.

In each case, I don't think it makes sense to try to find the filter with less worse time domain behavior.
Either you agree that it works like this, or you have to obsess a little bit in repeating the measurements as long as you accidentally alter the responsible variable (assuming it exists in a heuristic approach...?).
But it is quite illogical after paying for not having to stress yourself with the alignment of the sub(s)...
IMO

 

[Davide]

Following all comments here, I tried 10 different filters in DLBC to see if one would stand out from the crowd. It is hard with my knowledge to decide. GD is true only for one mic position meaurement? And for one channel at a time? DLBC filter is calculated on an average of 13 measurements. Why to prioritize one GD measure against 13?
How to choose between this filter results for the Front left channel

View attachment 409715

and this other filter?
View attachment 409716

PS. I got an idea but I can't test it at the moment.
Can you try to create a filter with only one measuring point? (central)
I know it can be done because I did it once, Dirac only tells you that there are few measurements, but it still makes you create the filter.
Or use the small 9-point measurement scheme and never move the microphone, which perhaps creates fewer problems for the algorithm.
This could show how the algorithm behaves in a situation with fewer time domain variables, and exclude that this GD is the result of the minimization of the Mean Square Response on several points.

 

[RenPa]

In each case, I don't think it makes sense to try to find the filter with less worse time domain behavior.
Either you agree that it works like this, or you have to obsess a little bit in repeating the measurements as long as you accidentally alter the responsible variable.
But it is quite illogical after paying for not having to stress yourself with the alignment of the sub(s)...
IMO

Sooooo true!

 

[RenPa]

PS. I got an idea but I can't test it at the moment.
Can you try to create a filter with only one measuring point? (central)
I know it can be done because I did it once, Dirac only tells you that there are few measurements, but it still makes you create the filter.
Or use the small 9-point measurement scheme and never move the microphone, which perhaps creates fewer problems for the algorithm.
This could show how the algorithm behaves in a situation with fewer time domain variables, and exclude that this GD is the result of the minimization of the Mean Square Response on several points.

Of course. I always create 3 liveprojects while measuring : MLP/ MLP+4/ and MLP +12.
I made 8 calculations with MLP only measurement in DLBC. All predicted are the same I usually got since 2 years. Some come back twice but suffice to say that every new calculation = a new filter response! After a while though, maybe I see 3 recuring patterns.

So correct me if I am wrong but it seems that only the first MLP measurement dictates the GD ( or lack of) calculation?

 

[Davide]

Of course. I always create 3 liveprojects while measuring : MLP/ MLP+4/ and MLP +12.
I made 8 calculations with MLP only measurement in DLBC. All predicted are the same I usually got since 2 years. Some come back twice but suffice to say that every new calculation = a new filter response! After a while though, maybe I see 3 recuring patterns.
View attachment 409931View attachment 409932View attachment 409933
View attachment 409934View attachment 409935View attachment 409936


View attachment 409937

So correct me if I am wrong but it seems that only the first MLP measurement dictates the GD ( or lack of) calculation?

Sorry, I wasn't precise in the previous post. I intended to create a single-point based filter, and then measure the GD and the spectrogram with REW.
Do you still get that anomaly?

 

[RenPa]

Sorry, I wasn't precise in the previous post. I intended to create a single-point based filter, and then measure the GD and the spectrogram with REW.
Do you still get that anomaly?

Not the same!
Do you mean to calculate only one filter should be enough? or Do I need to load 2-3 filters with different predicted response calculated with the same mlp measurement ?

 

[RenPa]

With firmware 3.11.00 version I have those one position mlp measurements ( I work with 3.4.4 usually but I did those measurements for Dirac to show that this latest didn't fixed DLBC bugs with Arcam)
Front left GD

Front left wavelet

 

[Davide]

Not the same!
Do you mean to calculate only one filter should be enough? or Do I need to load 2-3 filters with different predicted response calculated with the same mlp measurement ?

I'm afraid we don't understand each other.
My idea is this:
- create a new Dirac Live project
- select the "narrow" measurement scheme in Dirac Live, corresponding to a cube with 9 measurement points
- put the microphone in the MLP and capture all 9 measurements at that point, without moving it
- proceed to create the filter with the target and the crossover that Dirac proposes by default and load it into the processor
- without moving the microphone, measure the system response with REW and post the GD and spectrogram graph that you obtain
If at this point the GD trend is still anomalous (worse than original system) I would exclude the hypothesis that it is the (or one of the) best solution to minimize the mean square freq deviation between points and I'm starting to believe more in the other previous hypotheses (compromise to maintain low latency or time domain controls calibrated to questionable audibility levels or bug/inefficacy in the algorithm when particular acoustic conditions arise).
But the fact remains that we can do nothing about it, except complain if the final (corrected) sound is worse than the initial one (not corrected). But this will be really hard... I suspect no matter how much excess GD is introduced (as long as we're talking about 200ms), an improved FR will always be preferable.
For this reason I have the doubt that this behavior is somehow approved or a compromise rather than accidental (even if not always present, therefore dependent on the specific system).
Let's await feedback from technical support.

 

[RenPa]

Agree to await. I would prfer to do the exercise if and when Dirac will finally fixes the bugs in Arcam we have since 3.4.4 because I suspect that 3.11 may do a better job.

 

[LIΟN]

I suspect no matter how much excess GD is introduced (as long as we're talking about 200ms), an improved FR will always be preferable.

You mean exactly "200ms"?

 

[LIΟN]

But this will be really hard... I suspect no matter how much excess GD is introduced (as long as we're talking about 200ms), an improved FR will always be preferable.

You can listen to it through IEM or headphones. I asked for permission from the original owner of the measurement and converted the mdat file to IR, then looped a part of the song’s track. Yes, this is different from actually listening to it, as it’s not a true binaural recording (lacking HRTF), so ignore the overall tonal balance and high frequencies. Of course, there are changes in FR, and I could correct for those two differences, but if it’s the intended result from Dirac, it should not be altered. So, I’ve only briefly compared the filter settings as they are. Please listen and let me know what you think.

 

[LionIT]

I listened to the track and the difference is obvious and in favor of DLBM (presumably for lower GD).
But I'm not even sure that IR on a single point is representative of what Dirac does, as it works to optimize a volume.
In theory, BM applies crossover filters and corrects the response for individual units at multiple points, without considering the interaction between them, while BC considers this as well.
And on several points the interaction is really complex, and difficult to represent mathematically. In fact, for this reason it seems to be understood that Dirac uses a heuristic approach (it seems... the information is not so clear and exhaustive).
I assume that BC is designed to find an acceptable solution in this sense, favoring other factors than GD (ie. deviations between points, latency, ...).
Now, the fact that BM is acoustically preferable (in that case and for one point) does not necessarily mean that BC is fault.
Even if honestly, from a certain point of view, I hope so, so there is more possibility that work is done to correct it
Unfortunately with the music I don't like the bass I get with BC.
With films, however, the GD increasing as the frequency decreases, gives a certain presence. It is as if the individual frequencies are perceived better.
But technically I disagree with it.

 

[LIΟN]

I'm not even sure that IR on a single point is representative of what Dirac does,

Yes, exactly.
A single-point IR, especially one that isn’t a binaural recording, cannot represent it accurately.
That’s why I mentioned it in advance.
However, I’m curious about that.
Ultimately, if there is an optimal setup achievable in the room, it should be similar.
After all, the room’s response isn’t something we can selectively choose.

For example, this is a different topic, but:
The image above shows an ideal simulation decay for a typical multi-subwoofer setup,
and the image below shows an ideal simulation decay for a DBA configuration.
Both share the same room size and absorption rate. ---> Like this, I see that if there is an optimal setting in that space in any way, the results should be similar. Even if it's DLBM/DLBC.

I believe the greatest advantage of DLBC lies in its ability to provide more automated choices for the user’s convenience.
If the automated selection is effective and there are no issues in DLBM, the results of the two shouldn’t differ significantly.
(ART, of course, would be a separate discussion.)
Of course, the results may vary slightly depending on the processing algorithm.

that's why I uploaded a short loopback recording video of DLBM and DLBC.
The discussion about the difference between DLBM and DLBC is ongoing, but couldn't we think of it from the opposite perspective? If we had directly measured and set the delay, polarity, and EQ designed by DLBC, would we have used those values?

 

[LionIT]

If the automated selection is effective and there are no issues in DLBM, the results of the two shouldn’t differ significantly.

I understand what you mean, but I don't entirely agree, in the sense that BM solves the problem of FR in the time domain on several points for one unit (driver), while BC must do the same thing for two or more units summed (main + multiple subs).
This is mathematically a nightmare and obviously does not have a single solution, but more than one depending on criterion for which it is resolved.
It is a bit like the often reiterated concept that a DSP cannot effectively linearize an audio system in a non-infinitesimal point because there are variables that cannot be managed effectively, and the more it corrects at one point the worse another.
So the more variables you introduce, the more the speech is amplified.
So I don't think it makes sense to compare BC and BM, technically.
Obviously this does not take away the speech of the lower GD preferability (but also one point Vs multiple point correction preferability).

 

[LIΟN]

I understand what you mean, but I don't entirely agree, in the sense that BM solves the problem of FR in the time domain on several points for one unit (driver), while BC must do the same thing for two or more units summed (main + multiple subs).
This is mathematically a nightmare and obviously does not have a single solution, but more than one depending on criterion for which it is resolved.
It is often reiterated that a DSP cannot effectively linearize an audio system in a non-infinitesimal point because there are variables that cannot be managed effectively, and the more it corrects at one point the worse another.
So I don't think it makes sense to compare BC and BM, technically.
Obviously this does not take away the speech of the high GD.

I agree with you. No equivalent technical comparison, and my story is that if there is a response that can be best set in the same space, same speaker, and sub position, shouldn't the two results be similar? I think it was about and my nuances would have been conveyed

I've tried to reverse the GD increase in my own way, but I don't know as I keep saying. I don't think I can know anything without Dirac's answer, as other users say

 

[LionIT]

I agree with you. No equivalent technical comparison, and my story is that if there is a response that can be best set in the same space, same speaker, and sub position, shouldn't the two results be similar? I think it was about and my nuances would have been conveyed

I've tried to reverse the GD increase in my own way, but I don't know as I keep saying. I don't think I can know anything without Dirac's answer, as other users say

Yes, in practice, the ideal acoustic solutions should not differ much from each other ... but it depends a lot on the variability of the data I think.

However I hope that ART comes out soon. Early measurements show impressive time domain capacity.

 

[RenPa]

And on several points the interaction is really complex, and difficult to represent mathematically. In fact, for this reason it seems to be understood that Dirac uses a heuristic approach (it seems... the information is not so clear and exhaustive).
I assume that BC is designed to find an acceptable solution in this sense, favoring other factors than GD (ie. deviations between points, latency, ...).

I understand that DLBC algoritm is a learning machine. What I can't understand is why hitting calculate many times gives always different solutions (predicted filter SPL. After, say 5-6 tries, DLBC should be able to give the best one because the measurements are the same, the room is the same, the ETC are the same etc. It is like DLBC says: OK you don't like this one, try this one. OH not this new one, OK there's an other one etc until the end of Times!

 

[Davide]

You mean exactly "200ms"?

I mean a flattened FR with GD of that order could be preferable to uncorrected FR with lower GD.
With th same FR a lower GD will perhaps be preferable. Certainly more correct.

 

[Davide]

I understand that DLBC algoritm is a learning machine. What I can't understand is why hitting calculate many times gives always different solutions (predicted filter SPL. After, say 5-6 tries, DLBC should be able to give the best one because the measurements are the same, the room is the same, the ETC are the same etc. It is like DLBC says: OK you don't like this one, try this one. OH not this new one, OK there's an other one etc until the end of Times!

I believe it is due to what they called a genetic algorithm. Now I can't find where I read it.
Something heuristic to find more easily a solution to a complex problem. Which in the case of BC probably alternates between two or three responses due to how it is configured.
But I'm not an expert on that ...

EDIT: found a reference in this post

 

[RenPa]

And yesterday I tried a new filter with an othepredicted response confirmed by REW FR response which even better than any previous one. And, as usual, it is impossible to find why it is better.
Finally the conclusion: DLBC with 3.4.4 version and Arcam is a lotery...