Keith_W DSP system

[OCA]

With stereo it should be slightly behind your left or right. can’t really remember now. That’s how it sounded during my stereo days with the Harbeth speakers. I tried looking for the article describing how each tracks of the CD should project the sound. With XTC, I think it still about the same. Will listen again this weekend.

I should have written "...moves to further back with it."

 

[5-pot-fan]

Hi,
Thank you for this thread - illuminating, encouraging, entertaining and just a bit mind-blowing!
In a recent post you mentioned an effect I have been trying to pin down for while. You said "the center image is unstable and shifts depending on whether he is singing a high or a low note". Can you or anyone give a name to this? It is probably well-known and I have just missed it, but any help appreciated.
@OCA responded saying "In my view, the main reason for that is phase response differences between speakers around 60Hz-1000Hz". I can grasp the idea but any further observations would be nice, if available.
In my small system I hear the effect only occasionally and it is with electric guitar passages, but I will now search out some more operatic pieces and see/hear what I can.
It was also good to learn that your wife approves of the room changes!
Please keep this thread running, at least while you nail down all the sonic changes.

 

[OCA]

Hi,
Thank you for this thread - illuminating, encouraging, entertaining and just a bit mind-blowing!
In a recent post you mentioned an effect I have been trying to pin down for while. You said "the center image is unstable and shifts depending on whether he is singing a high or a low note". Can you or anyone give a name to this? It is probably well-known and I have just missed it, but any help appreciated.
@OCA responded saying "In my view, the main reason for that is phase response differences between speakers around 60Hz-1000Hz". I can grasp the idea but any further observations would be nice, if available.
In my small system I hear the effect only occasionally and it is with electric guitar passages, but I will now search out some more operatic pieces and see/hear what I can.
It was also good to learn that your wife approves of the room changes!
Please keep this thread running, at least while you nail down all the sonic changes.

I have recently posted various test tracks for that:

Critical (Best) Music Tracks for Speaker and Room EQ Testing

https://www.audiosciencereview.com/forum/index.php?threads/critical-best-music-tracks-for-speaker-and-room-eq-testing.6/

www.audiosciencereview.com

 

[Keith_W]

Neither the sofa nor the mic position would make such a difference m8. It looks like it's around 50-60Hz. I don't know what frequency is the AC down under but it might be a problem with your AC lines maybe? I had similar major jumps or dips from one measurement to the next due to phase shifts around that region in a previous apartment with problematic wiring. I have also "rarely" seen such differences between left and right speakers in measurements sent to me from some quite randomly treated rooms.

That same area (60Hz) switches back and forth also in the spectrograms with the doors btw

After I read your post I thought "pffft". Then I remembered that I have an AC monitor plugged in:

WTF! 200V, 32Hz!?!?!?!?

In one of my earlier posts I mentioned a thunderstorm which took out power infrastructure in Melbourne. I do not expect 200V, it should be 240V, 50Hz! Unfortunately, this is the only monitor I have, so I can't verify if it is malfunctioning or not.

Well, I do have a multimeter but I am not very skilled at using it, and I am not confident of pushing the probes into the power socket. If you know where I should turn the switches to, I would be grateful:

 

[sweetsounds]

WTF! 200V, 32Hz!?!?!?!?

In one of my earlier posts I mentioned a thunderstorm which took out power infrastructure in Melbourne. I do not expect 200V, it should be 240V, 50Hz! Unfortunately, this is the only monitor I have, so I can't verify if it is malfunctioning or not.

Looks like a flawed meter, here is how to use the multimeter:

Select the V with the wave on your multimeter. Your device doesn't have a FREQ mode to measure frequency.

Not a difficult task, but please follow the guidelines when working with voltage <1000V: e.g. don't touch the metal tips, don't work in a cramped space (e.g. you need to be able to pull back your hands).

 

[Keith_W]

Well it seems as if the wall monitor is kaput. The wall monitor is still reading 200V while the multimeter says 240V. Of course the Hz could still be wrong, but I have no way to test that.

I will have a think about why there is such a massive discrepancy between the repeat measurement today and the verification measurement from a few days ago. I will remove the sofa and repeat.

 

[Keith_W]

In a recent post you mentioned an effect I have been trying to pin down for while. You said "the center image is unstable and shifts depending on whether he is singing a high or a low note". Can you or anyone give a name to this? It is probably well-known and I have just missed it, but any help appreciated.
@OCA responded saying "In my view, the main reason for that is phase response differences between speakers around 60Hz-1000Hz". I can grasp the idea but any further observations would be nice, if available.

Thank you, I am glad that you appreciate this thread. Umm, this question you asked is a little bit difficult to explain, but I will try.

Firstly, there are a lot of people who think that phase shifts are not audible and you should not worry about them. However, changes in relative phase ("out of phase") are definitely audible. Have a listen to this:

What is happening is that the left speaker is pushing out, while the right speaker is pushing in. While the "in phase" version sounds outside, centered, and in front of you, the "out of phase" version sounds like it is inside your head. You can get the same effect by reversing the polarity of one of your speakers. If you reverse both, it will sound exactly the same as when it was not reversed except at the crossover point, where you may get strange summation/cancellation effects.

A reversal in absolute polarity of one speaker means that one speaker is 180deg out of phase with the other, which produces that effect. Less phase shift between L/R speaker has the same effect, but it progressively becomes more subtle. What you hear are shifts in the center image which varies with frequency.

What happens in the room is a soup of phase. Hidden somewhere in there is the minimum phase signal. You can read more about it here. The min phase signal has the important property that inversion of this signal corrects for phase and frequency response at the same time. But in order to do that, you need to get rid of everything else which is not minimum phase, known as excess phase. The calculation involved is somewhat complicated and explained in that article I linked. Fortunately, Acourate does this for you and automatically subtracts it. If you don't subtract it, or you do it improperly, you will add to the phase problems rather than fixing it.

I think this is what OCA is suggesting, and if it is not - he can clarify it himself.

 

[AudioJester]

Looks like the output of the midrange drivers has dropped. Is it an amp/output issue to those drivers?

 

[ernestcarl]

I guess if there is one good thing to say about the spectro, the decay time of all frequencies seems to be constant. But this system is really turning out to be a head scratcher. I can certainly hear the difference the room treatment makes, but I can't seem to measure it.

You should be able to see a difference in the spectrogram (more so the wavelet) with even a single or a few small 12x12 inch foam acoustic absorber panels when placed at exactly at a first early reflection point that happens to come back towards the mic position -- you just have to zoom-in really, really close in time and play around with the parameters (e.g. 15 ms). *Then, again, my actual experiments and experience here is coming from my (already) dry room to begin with -- where the spectrograms are probably much easier to read...

 

[ernestcarl]

I suggest you try to use and keep readily on-hand a known "flat" active speaker (e.g. KH80) for comparison purposes with your own system as a sort of "secondary check". It might save you time figuring things out whenever weird inconsistencies in your measurements happen.

 

[Keith_W]

Looks like the output of the midrange drivers has dropped. Is it an amp/output issue to those drivers?

Don't know. Seems highly unlikely given they are monoblocks. There is something going on, I will find out later. I plan to do some investigating today.

I suggest you try to use and keep readily on-hand a known "flat" active speaker (e.g. KH80) for comparison purposes with your own system as a sort of "secondary check". It might save you time figuring things out whenever weird inconsistencies in your measurements happen.

Great, more money to spend!!!

 

[ernestcarl]

Great, more money to spend!!!

Just a suggestion to keep in mind. Every time a measurement anomaly comes up -- which seems to be fairly regular in your case -- what other static, reliable reference do you have to compare against? Doesn't even have to be a Neumann studio monitor at all. There are many active speakers that are much cheaper that have a reasonably reference "flat" response.

 

[Keith_W]

Well, I figured out what caused the issue with the funny verification measurement. I was using the wrong (*)*(&))&)^(*& filter! Rather than using the verification filter (i.e. the one I make after doing all the room correction), I was using the "initial" filter. I remade the filter and did the correction, and it looks exactly the same as the verification measurement. That big X on the floor helps with consistent microphone placement.

 

[5-pot-fan]

Thank you, I am glad that you appreciate this thread. Umm, this question you asked is a little bit difficult to explain, but I will try.

Firstly, there are a lot of people who think that phase shifts are not audible and you should not worry about them. However, changes in relative phase ("out of phase") are definitely audible. Have a listen to this:

What is happening is that the left speaker is pushing out, while the right speaker is pushing in. While the "in phase" version sounds outside, centered, and in front of you, the "out of phase" version sounds like it is inside your head. You can get the same effect by reversing the polarity of one of your speakers. If you reverse both, it will sound exactly the same as when it was not reversed except at the crossover point, where you may get strange summation/cancellation effects.

A reversal in absolute polarity of one speaker means that one speaker is 180deg out of phase with the other, which produces that effect. Less phase shift between L/R speaker has the same effect, but it progressively becomes more subtle. What you hear are shifts in the center image which varies with frequency.

What happens in the room is a soup of phase. Hidden somewhere in there is the minimum phase signal. You can read more about it here. The min phase signal has the important property that inversion of this signal corrects for phase and frequency response at the same time. But in order to do that, you need to get rid of everything else which is not minimum phase, known as excess phase. The calculation involved is somewhat complicated and explained in that article I linked. Fortunately, Acourate does this for you and automatically subtracts it. If you don't subtract it, or you do it improperly, you will add to the phase problems rather than fixing it.

I think this is what OCA is suggesting, and if it is not - he can clarify it himself.

Thank you. I will read through your links and what you and @OCA have said and will return with queries if appropriate to this thread.

 

[Keith_W]

I had a session with @UliBru last night, and he made me a new filter. As usual, I learnt a lot. TLDR: his sounds better, but mine measures better. Read on to find out why.

First, let's take a look at the verification measurements of the two filters:

These are verification measurements of @UliBru's filters (red/green, upper), and mine (blue/brown, lower). These used almost the same target curve. His curve has a notch at 4.5kHz to avoid overcorrecting the tweeter, whereas mine does not. Other than that, you can see that they both follow very similar shape. Apart from the target curve, the only differences between these filters is the driver linearization procedure (his vs. mine), and workflow settings for room correction. The XO points and slopes are exactly the same.

I had a friend come over to listen to the two filters. The difference is more obvious with his choice of music vs. mine (he listens to pop, I listen to classical). Dr. Uli's correction has a certain openness and spaciousness, and even the height of the soundstage is different (his is lower, mine is higher). The tone is also different, his correction seems to be more neutral, whereas mine sounds a bit darker. My filters sound really tight, whereas his filters sound relaxed. As my friend said, mine sounds boxy, his sounds open. Instruments with my correction seems to be tightly located in a box in front of you, whereas the placement of his instruments seems more believable. I am sorry, I am not so good at describing the sound, but be assured that the differences are real!

If we overlap the verification measurements and only show the left speaker, you can see the difference between his correction (red) and mine (brown). OK, so we know that a difference in the target curve will produce different tonality. So that part is explained. But why does it sound more spacious? Or, as my friend put it, my correction sounds more "boxy". So let's look at some other curves.

Comparison of his correction (red) vs. mine (brown), with phase extraction to show the min phase version. Look at the phase angle. He did not linearize the phase of the tweeter with a reverse AP filter, whereas I did. This is why mine shows almost flat phase response from 200Hz up to 24kHz, whereas the red curve shows some phase rotation at the XO point to the tweeter.

On a graph, my correction certainly looks prettier. But his sounds better. I can't wrap my head around WHY but Uli said that the behaviour of my tweeter is odd, and correction might make it worse, so it is better to leave it alone.

This is the unwrapped non-minphase version, i.e. without phase extraction. You can see the difference between his correction and mine.

Could the difference be due to group delay? Not really, my version (in brown) has better GD than his.

One notable difference is that his step response looks much better. He spent a lot more effort on time alignment than I did. Above in red/green is his step response.

And this is mine. Both graphs are displayed in the same scale. Mine has more pre-ringing and the alignment of all the pulses isn't as good as his.

As mentioned, there were a few differences in his workflow vs. mine:

1. He did not linearize the tweeter, because he thought it behaved oddly. Mine was linearized until it was absolutely flat.
2. He did not linearize the subwoofer, because he thought that an in-room measurement of the sub would be contaminated by the room, making correction impossible. He feels the same about lower woofer frequencies, but he linearized the woofers anyway.
3. He does his driver linearization with the mic 1m from the drivers. When I do mine, the mic is almost touching the driver. I forgot to ask him why he does this. It may be to capture the baffle step response, I don't know.
4. When he does the excess phase correction, he repeats iterations whilst watching for group delay. This messes up the step response and introduces pre-ringing, which he then compensates for with Acourate's pre-ringing compensator. When I do mine, I ONLY watch the step response. Despite this difference in workflow, the results are opposite - my GD looks better than his, but my step response looks worse than his.
5. He uses three different time alignment procedures to align the subwoofer and he cross-checks them with each other. I only use one, and my only check is my verification measurement where I look at the impulse and step response.
6. I use 65536 taps. He prefers 131k or 262k taps, but for me the latency is excessive. He offered, I declined. For those who don't know, latency can be calculated by: (N-1) / (2Fs), where N = number of taps, and Fs = sampling frequency. So if we assume we have a perfect convolver, the latency for 65536 taps at 48kHz is 0.682 seconds. For 131k taps, it is 1.37 seconds, and for 262k taps it is 2.73 seconds. The advantage of going with higher taps is finer correction; the resolution of 64k taps at 48kHz is (48000/65536) = 0.73Hz, 128k taps is half of that (0.37Hz), and 256k is half that again (0.18Hz). I personally don't think I could ever hear the difference between a 0.73Hz correction and 0.18Hz correction, but I would certainly notice the increased latency which doubles every time.
7. He seems to prioritize volume over correction, whereas I tend to sacrifice volume for correction. As seen above, he even notched out the tweeter XO point to gain more volume. As a result, his filter is louder than mine and noticeably lumpier. Don't worry, I volume levelled the two filters before listening (Acourate Convolver has a volume levelling feature).

Because I also use Acourate Convolver, he made some modifications to my convolver settings (specifically AcourateFLOW and AIR). Separate discussion to follow.

So in summary: my corrections generate prettier looking graphs. But his sounds better. And it's not even close. I suspect it comes down to knowledge of what to correct and what to leave alone that has created this discrepancy. I look at online graphs, I know what perfect looks like, and I try to generate a perfect looking graph from imperfect measurements. I succeed, but I end up correcting for things I shouldn't and it sounds worse.

To be honest, right now I am feeling a little bit lost, as if something has been taken away from me. There is obviously a large gap in knowledge and learning between him and me, only this time I don't know where to look to improve my learning. I will have to hit the books again to look for answers, there must be something I skimmed over.

 

[ernestcarl]

The magnitude levels in Ulli's correction above that 4.5 kHz notch needs some further explanation. I'm sure he's got a good reason to keep it slightly off.

The differences are not really that striking at first glance, but you may be over-correcting to fit a certain target expectation far too perfectly e.g. too tight tolerances and step response and GD looks "compressed" in time relative to Uli's -- "better measuring" may only be apparent on the surface. A method of visualizing how much of a drastic change occurred pre- and post-EQ would be by carefully aligning and overlaying the uncorrected vs corrected measurements across different plot views -- it might be easier to manipulate certain graphs for the stuff I'm personally interested in seeing close-up with REW. Hmmmn... although, having your Acourate measurements "translate" into REW is going to require additional work, unfortunately. And, you have a four-way system so it will be quite tricky to show this optimally using only single graphs. Splitting it into one or two sections at a time would make for easier visualization.

 

[ernestcarl]

So if we assume we have a perfect convolver, the latency for 65536 taps at 48kHz is 0.682 seconds. For 131k taps, it is 1.37 seconds, and for 262k taps it is 2.73 seconds. The advantage of going with higher taps is finer correction; the resolution of 64k taps at 48kHz is (48000/65536) = 0.73Hz, 128k taps is half of that (0.37Hz), and 256k is half that again (0.18Hz).

Why does it have to be a "perfect convolver"? Impulse centering is arbitrary -- the delay can be set to be near 0ms with high tap count if we set aside phase correction, for example. Even a partial correction could be adequate for the task if filter economy and short latencies were more prioritized.

 

[Keith_W]

The magnitude levels in Ulli's correction above that 4.5 kHz notch needs some further explanation. I'm sure he's got a good reason to keep it slightly off.

I am pretty sure it was a measurement artefact and isn't deliberate. This is what the test convolution looked like:

Those tweeters are auto-sensing, and they turn themselves off when there is no signal. They automatically wake up with signal, but they take some time to stabilize. There is audible hissing and crackling for a few minutes, and then they go nice and quiet. During this time the volume may be a bit variable. I can hear it with music, when I turn the system on to listen, it sssssssssssoundsssss ssssssssssibilant for a few minutes then it settles down. I had just woken the tweeters up to make the verification measurement, which is probably why there is a discrepancy. I just made a video, not sure if it will work: click to view.

The differences are not really that striking at first glance, but you may be over-correcting to fit a certain target expectation far too perfectly e.g. too tight tolerances and step response and GD looks "compressed" in time relative to Uli's -- "better measuring" may only be apparent on the surface. A method of visualizing how much of a drastic change occurred pre- and post-EQ would be by carefully aligning and overlaying the uncorrected vs corrected measurements across different plot views -- it might be easier to manipulate certain graphs for the stuff I'm personally interested in seeing close-up with REW. Hmmmn... although, having your Acourate measurements "translate" into REW is going to require additional work, unfortunately. And, you have a four-way system so it will be quite tricky to show this optimally using only single graphs. Splitting it into one or two sections at a time would make for easier visualization.

The reason I wrote down all the differences between his method and mine is because I am kind of using this thread as a notebook for lessons that I have learnt. Or in this case, a few things that I am going to have to unlearn. It really pains me to say this, but I am starting to realize that perfect looking measurements does not equal perfect sound. NOT if the "perfect looking measurements" are valid only for one point in space and based on incorrectly taken initial measurements with faulty assumptions made during the correction. What bothers me is that I don't know what it is that I am ignorant of, because at least if I knew where my ignorance was, I could go and fix it.

Don't get me wrong, the filters that I made sound pretty good. It's just that Uli's are better. A few pages back you can see that I said something along the lines of "it could always sound better, just that I don't know it yet". Well, now I know that it really can sound better. And I think it can sound better still, if I take Uli's suggestion to move the tweeter XO point up to 6kHz to avoid that 4.5kHz notching problem. This is something I will have to seriously investigate.

Why does it have to be a "perfect convolver"? Impulse centering is arbitrary -- the delay can be set to be near 0ms with high tap count if we set aside phase correction, for example. Even a partial correction could be adequate for the task if filter economy and short latencies were more prioritized.

Hmm, so you are saying that a 64k tap filter is the same as a 128k tap filter if they both have exactly the same settings for phase correction, delays, etc? I have to say I am a bit skeptical, but there is only one way to find out.

 

[Keith_W]

@ernestcarl if you want to look at the difference between the initial measurement and corrected measurement, here it is:

The volume on the subwoofer is maxed out to give plenty of scope for cutting.

 

[ernestcarl]

Hmm, so you are saying that a 64k tap filter is the same as a 128k tap filter if they both have exactly the same settings for phase correction, delays, etc? I have to say I am a bit skeptical, but there is only one way to find out.

a partial correction could be adequate for the task if filter economy and short latencies were more prioritized.

One does not necessarily have to lump everything into the FIRs -- the task can be split.


Ex. of my center spkr+sub at the couch (I know optimal DSP corrections are very setup dependent):

We could make a zero FIR latency version from the raw measurements with no additional phase correction applied. Since the room's reflection-decay is already quite short, the benefit in the low end here is relatively small. The bass below 140 Hz is deliberately LS boosted in the version without FIR filter optimization around the xo bass region a tiny bit to compensate since the impulse delay makes it sound psychoacoustically a little less loud overall.


But, yeah, I know there is a specific workflow with Acourate...