Finally, as promised (but similar to a long line of unmet promises), here is the report on the living room system time alignment and adjustment near the beginning of June 2019 when I finally turned REW onto the living room system in order to do a measured time alignment which hadn't systematically been done since about 2014...something like forever for this ever changing system.
In May I obtained a Focusrite 2i4 interface so I could do full loopback with REW. I had been unable to get my historic Emu 0404 to work on Windows 10, so I just decided to get the cheap interface everyone uses now (and which apparently still works on Windows 10, with the status of other interfaces such as the Lynx Hilo uncertain). I already had a suitable calibrated mike I bought specifically for REW a few years ago. You cannot use a USB mike during loopback because you must use the same audio interface for both input and output, and a USB mike only has an input interface.
When you do full loopback, in theory, you can simply have REW compute the delay time from the loopback electrical output to the measured acoustical stimulus. Then, in theory, you can simply dial in the required compensation into your digital processing units.
Using a loopback is also preferred for getting a more accurate computed step response. REW uses a stimulus which is a sine wave sweep from a low frequency to a high frequency. The actual response to that is hard to interpret for phase response by eye. So then REW analyzes the response to this sweep stimulus using FFT, and then it can compute what the response would be either to a dirac type impuse (the "Impulse" response), and a step type stimulus (the "Step" response). Because it is the simplest possible stimulus, and arguably most informative, speaker designers and reviewers mostly evaluate the step response. It should always be remembered, however, that REW did not actually measure this stimulus as such, it merely computed that, given the way the unit under test responded to the sweep, this is the way it would have responsed to an actual step. The computed step response will have artifacts. But if you were measuring an actual step, it would be highly random because of noise. The swept sine stimulus, having more broadspectrum (across desired band) energy increases the S/N over what you could measure with an actual Step stimulus.
I had long done such compensation simply based on the physical distances of the speakers as seen from the listening position. Roughly speaking, the supertweeters are about 1 foot back from the panels, and the subs are about 2 feet back. Assuming a foot of distance to have about 1 ms delay (I'll use that approximation a lot in this post), I could then compensate by having NO DELAY (0 ms) dialed in to the subwoofer crossover unit, 1 ms delay dialed in to the supertweeter crossover unit, and 2ms dialed in for the panels. In practice, I intentionally add 10ms to these numbers to facilitate more on-the-fly experiments and adjustments, since you can't further reduce a delay already at 0 ms. So, a typical adjustment based on distances might be: 10ms delay on panels, 9ms delay on supertweeters, and 8ms delay on the subs. (In practice, I tried to make the numbers accurate to 0.1ms by measuring to the nearest inch.)
But this is all approximations until you actually do acoustical measurements, especially when all speakers are nothing like point sources. Real acoustical measurement, not distance approximation, is what has to be done to do the time alignment correctly. But it's so difficult, and fraught with problems, I haven't actually done it very much. In principle, I should do a new time alignment after any adjustment of any kind.
It was failure to do actual acoustical measurements (which many subjectophiles and speaker amateurs HAVE NEVER DONE) that led to a huge mistake which I continued for several years! I started using different DACs on each way, without realizing that the DACs themselves have delay differences, and these differences change at each different sampling rate. Once I discovered this issue, I quickly gave up on dialing in new compensation when the sampling rate changed. Too much work. But that was when I decided to add 10ms extra delay to each path. However, it turned out that some DACs could add far more than 10ms delay for some sampling rates. I ultimately decided the best (?) way to handle that variance is to eliminate it by using identical DACs on each way. I went with inexpensive but decent measuring and pure sounding Emotive Stealth DC-1's (now discontinued) because they had all the needed features: AES digital input, balanced and unbalanced outputs, exact level adjustment. Audio Science Review puts this DAC down (somewhat unfairly IMO) however if I had to make the choice now, I'd go with the similarly inexpensive DAC they recommend, which has even lower distortion and noise.
This time around, my plan on simply dialing in the REW computed delay for each way was swiftly defeated by an apparent fault with the way REW computes delay for my subwoofers. I have written a long post about this previously. I first noticed that the delay times being computed by REW for the subwoofers were about 3ms longer than I had expected. Then, looking at the computed step response, it appeared that REW was computing the zero point as the beginning of the second hump in the impulse response. This appears to be an error in REW--while the second hump is slightly larger than the first, it is nearly the same height, and does not appear to be a digital or noise artifact (though there are lots of those, so it's not certain either). It may have something to do with peculiarities in my system--the rear wall reflection is my current best theory. This is an example of the computed step response for the Right subwoofer. The red line shows where REW believes the actual step response is beginning:
It shocked me to see that I got this same result even with No Crossover, which I finally realized I could safely do simply by restricting the range of the sweep to below 300 Hz. In tests a few years back I decided it might not be safe to send a full range signal--especially a dirac--to the subwoofer with no internal or external crossover being applied. So I feared sending audio to the sub without crossover. But this is safe when the input stimulus itself is restricted. Coincidentally REW defaults to a range below 250 Hz, making it safe to run on most subs without crossover, until further adjustments by the user. Here is the computed step response with No Crossover:
So, the smaller leading hump which REW's delay calculation ignores has nothing to do with the use of a crossover, which had been the leading theory until I took that measurement (and it's still hard to get out of my mind, because I remembered either lowpass or highpass was supposed to LEAD...but actually it's the highpass that's supposed to LEAD).
Along with the double hump is the extremely curious fact that the beginning of the first hump actually has less delay than the output of the panels, even though the subs are further back. The beginning of the second hump, where REW seems to think the step response begins, has about 3ms more delay than would be guess from the physical location of the speakers.
Now, REW seemed to give me two choices: go with the officially determined "delay time" which put the sub about 6 feet back from the panels instead of the approximate 3 foot distance, or go with the time of the beginning of the first hump, which put the subs ahead of the panels by about a foot (???)
I was so frustrated by this, I decided to look at something else: the notch in the panel response around 225 Hz which I had long assumed to result from rear reflection/cancellation. I discovered that moving the speakers further out from the wall could change this a lot, seeming to almost eliminate the problem, and I decided to optimize THIS before doing the actual time alignment. As I changing the Acoustat position, and looking only at the frequency response btw, I simply left the delay compensation as I had previously set it by distance approximation. Starting from the "original" position from a few years ago, which had the 225 Hz suckout, I was getting a step response like this:
Notice the little peak at just over the top of a large hump. That little peak is the beginning of the Acoustat step response, wheras the hump is the first hump of the subwoofer step response.
When I finally decided to move the Acoustats exactly 6 inches further out (for a total of about 39 inches from the wall) because it removed the suckout at 225Hz in the frequency response I was looking at, the step response looked like this:
You can see the little peak of the Acoustat step response moving forwards in time, as the Acoustats are being moved closer to the listening position, just about to the highest point of the first subwoofer hump. But even moved 6 inches forwards (which is about 5 inches closer to the listening position) the subwoofer response still seems to start first. Notice you can also see the little wiggle where the supertweeter is starting about 1 ms later than the Acoustats. At this point I had not yet moved the supertweeters forward to match the Acoustats, which I couldn't completely do anyway, and it appears their initial adjustment was slightly wrong also, but less than 0.5 ms as you can see in the earlier graph where the supertweeter start is a bit more subtle.
And now, for something completely different, here is what the Acoustat step response looks like just by itself (with the low and high pass crossovers I am using):
It pretty much starts with a spike upwards, as it should. Somehow the amplitude appears much greater than it does in combination with the woofer response, partly that's due to normalization and the curious nature of step response computation.
I measured and stared at the subwoofer step response for a long time, and finally made a wild guess about where the beginning of the woofer step should have been computed by REW. I then used the computed panel delay, and MY computed subwoofer delay, to figure out a correct delay compensation. Sadly I was not taking notes about any of this, it was just a guess, and I tried it, and this combined Acoustat and Sub alignment is about good as anything I've done since, which have only been minor excursions away from this point anyway:
The Acoustat step now merges with the woofer step so that the intial rise is a tall spike, as it should be. It still possibly looks like there is some preceding bass response, but that's down by at least 40dB, and is probably just background noise and artifact. I could not make the step look any better by delaying the bass much more than the above. I tried adding a bit more delay, and the step got worse:
OK, so whatever that was, it looks like too much subwoofer delay has pushed the initial Acousat spike down into the hump in some bass noise valley that precedes the actual bass response, which you can see kicking in a few ms after the initial spike with a blast.
Comparing these pictures begins to get subjective again, I admit, and I wish I had something more solid to latch on to. It did seem to me that I could try to maximize the length of the leading edge of the spike. That would indicate that all drive units are contributing maximally to that leading edge. Now, since REW always places the top of the spike at 0dB (btw, I have not yet calibrated the Focusrite interface itself) maximizing the length of the leading edge curiously means that the leading edge should "start" as low as possible. You can see that operating in the above two graphs. In the first graph, what I am calling the noise floor joins the leading spike at -46dB, with the top of the spike at -7dB, so I would count this as showing a leading edge of 39dB. In the second graph, the noise floor joins the leading spike at -30dB, so the leading edge of only 23dB, clearly inferior.
What isn't so easy to interpret, however, is the the negative spike excursion below the joining point means. I think it is mostly artifact, but also should be as low as possible indicating complete cancellation or some such. Anyway, as the delay is adjusted up and down the distance downward changes and also a "gap" seems to open up between the preceding bass-hump noisefloor and the leading edge of the step, making even more a subjective call each time. As I can't remember the actual adjustment for each measurement (my poor note taking on display again) I'm just going to show a bunch of them, in the order I took them, and where I stopped and why.
Trial 3 has leading edge 24dB, very poor.
Trial 4 has leading edge 30dB (but look at how low it goes also--actually to the bottom of the full scale not shown).
Trial 5 has leading edge 28dB, but deep spike again.
Trial 6 has leading edge 47dB and deep valley beforehand, best so far.
Trial 7 has leading edge 32dB, very poor.
Trial 8 has leading edge leading edge 31dB, very poor.
Stepping back in opposite direction now, Trial 9 has leading edge about 35dB.
51dB !!! This appears to be the winner, especially if you take the beginning of the initial spike at -58dB, then the initial spike is 51dB. The supertweeter is already being added in and that might account for some of the wiggle at the start--it appears pretty well aligned too. Alternatively, you could interpret the beginning of the spike as high as -48dB, in which case this is not the winner, but close.
With hand wave you could almost say this is almost looking like a pretty good step response, with a tall initial spike and pretty well filled in right after that.
This is clearly not as good, a step backwards, with the leading edge at no more than 40dB.
After the 11th measurement, I believe I went back to the preceding one, Alignment 10, as the best. I can't be sure of this, because I didn't take notes, so it might have been one of the other pretty good ones, such as The First Guess. But I'm pretty sure it was the 10th, because I could still remember the values for the 10th when I did the 11th, and the 11th was clearly a step backwards.
The curious delays in the DSP boxes to achieve the best alignment are:
Subwoofer: 6.3 ms delay
Supertweeter: 4.6 ms delay
Panels: 5.3 ms delay
So the subwoofers are delayed 1 ms MORE THAN than panels, though they are also about 2 feet further back. That is the second issue I don't understand.
I have already ruled out the lowpass crossover as causing this. It also probably not because of delays in the complex interface box of the Acoustats, because the delays required to align the supertweeters and the panels is about what one would expect based on distance.
This should not be confused with the first issue I don't understand, which is why REW is apparently computing the subwoofer delay time incorrectly. This is an entirely separate issue, as I did my alignment by the graphical means shown above, and NOT by using REW's delay estimates, especially for the subwoofer. However, there might be some connection I cannot yet fathom between the two. Or maybe not.
Now some might find it outlandish to write so much and not say a word about how it sounds. Of course it sounds wonderful! My "electrostatic" bass is now even more well integrated into the sound, without being the least bit less impactful. Bass lines are easier to follow. And in many other ways it sounds better overall.
But you should expect that I would feel this way, and it's probably not best to fully trust the audio judgements of the audio investigator, and probably not the measurements either.
Next month it could be different.
In May I obtained a Focusrite 2i4 interface so I could do full loopback with REW. I had been unable to get my historic Emu 0404 to work on Windows 10, so I just decided to get the cheap interface everyone uses now (and which apparently still works on Windows 10, with the status of other interfaces such as the Lynx Hilo uncertain). I already had a suitable calibrated mike I bought specifically for REW a few years ago. You cannot use a USB mike during loopback because you must use the same audio interface for both input and output, and a USB mike only has an input interface.
When you do full loopback, in theory, you can simply have REW compute the delay time from the loopback electrical output to the measured acoustical stimulus. Then, in theory, you can simply dial in the required compensation into your digital processing units.
Using a loopback is also preferred for getting a more accurate computed step response. REW uses a stimulus which is a sine wave sweep from a low frequency to a high frequency. The actual response to that is hard to interpret for phase response by eye. So then REW analyzes the response to this sweep stimulus using FFT, and then it can compute what the response would be either to a dirac type impuse (the "Impulse" response), and a step type stimulus (the "Step" response). Because it is the simplest possible stimulus, and arguably most informative, speaker designers and reviewers mostly evaluate the step response. It should always be remembered, however, that REW did not actually measure this stimulus as such, it merely computed that, given the way the unit under test responded to the sweep, this is the way it would have responsed to an actual step. The computed step response will have artifacts. But if you were measuring an actual step, it would be highly random because of noise. The swept sine stimulus, having more broadspectrum (across desired band) energy increases the S/N over what you could measure with an actual Step stimulus.
I had long done such compensation simply based on the physical distances of the speakers as seen from the listening position. Roughly speaking, the supertweeters are about 1 foot back from the panels, and the subs are about 2 feet back. Assuming a foot of distance to have about 1 ms delay (I'll use that approximation a lot in this post), I could then compensate by having NO DELAY (0 ms) dialed in to the subwoofer crossover unit, 1 ms delay dialed in to the supertweeter crossover unit, and 2ms dialed in for the panels. In practice, I intentionally add 10ms to these numbers to facilitate more on-the-fly experiments and adjustments, since you can't further reduce a delay already at 0 ms. So, a typical adjustment based on distances might be: 10ms delay on panels, 9ms delay on supertweeters, and 8ms delay on the subs. (In practice, I tried to make the numbers accurate to 0.1ms by measuring to the nearest inch.)
But this is all approximations until you actually do acoustical measurements, especially when all speakers are nothing like point sources. Real acoustical measurement, not distance approximation, is what has to be done to do the time alignment correctly. But it's so difficult, and fraught with problems, I haven't actually done it very much. In principle, I should do a new time alignment after any adjustment of any kind.
It was failure to do actual acoustical measurements (which many subjectophiles and speaker amateurs HAVE NEVER DONE) that led to a huge mistake which I continued for several years! I started using different DACs on each way, without realizing that the DACs themselves have delay differences, and these differences change at each different sampling rate. Once I discovered this issue, I quickly gave up on dialing in new compensation when the sampling rate changed. Too much work. But that was when I decided to add 10ms extra delay to each path. However, it turned out that some DACs could add far more than 10ms delay for some sampling rates. I ultimately decided the best (?) way to handle that variance is to eliminate it by using identical DACs on each way. I went with inexpensive but decent measuring and pure sounding Emotive Stealth DC-1's (now discontinued) because they had all the needed features: AES digital input, balanced and unbalanced outputs, exact level adjustment. Audio Science Review puts this DAC down (somewhat unfairly IMO) however if I had to make the choice now, I'd go with the similarly inexpensive DAC they recommend, which has even lower distortion and noise.
This time around, my plan on simply dialing in the REW computed delay for each way was swiftly defeated by an apparent fault with the way REW computes delay for my subwoofers. I have written a long post about this previously. I first noticed that the delay times being computed by REW for the subwoofers were about 3ms longer than I had expected. Then, looking at the computed step response, it appeared that REW was computing the zero point as the beginning of the second hump in the impulse response. This appears to be an error in REW--while the second hump is slightly larger than the first, it is nearly the same height, and does not appear to be a digital or noise artifact (though there are lots of those, so it's not certain either). It may have something to do with peculiarities in my system--the rear wall reflection is my current best theory. This is an example of the computed step response for the Right subwoofer. The red line shows where REW believes the actual step response is beginning:
It shocked me to see that I got this same result even with No Crossover, which I finally realized I could safely do simply by restricting the range of the sweep to below 300 Hz. In tests a few years back I decided it might not be safe to send a full range signal--especially a dirac--to the subwoofer with no internal or external crossover being applied. So I feared sending audio to the sub without crossover. But this is safe when the input stimulus itself is restricted. Coincidentally REW defaults to a range below 250 Hz, making it safe to run on most subs without crossover, until further adjustments by the user. Here is the computed step response with No Crossover:
So, the smaller leading hump which REW's delay calculation ignores has nothing to do with the use of a crossover, which had been the leading theory until I took that measurement (and it's still hard to get out of my mind, because I remembered either lowpass or highpass was supposed to LEAD...but actually it's the highpass that's supposed to LEAD).
Along with the double hump is the extremely curious fact that the beginning of the first hump actually has less delay than the output of the panels, even though the subs are further back. The beginning of the second hump, where REW seems to think the step response begins, has about 3ms more delay than would be guess from the physical location of the speakers.
Now, REW seemed to give me two choices: go with the officially determined "delay time" which put the sub about 6 feet back from the panels instead of the approximate 3 foot distance, or go with the time of the beginning of the first hump, which put the subs ahead of the panels by about a foot (???)
I was so frustrated by this, I decided to look at something else: the notch in the panel response around 225 Hz which I had long assumed to result from rear reflection/cancellation. I discovered that moving the speakers further out from the wall could change this a lot, seeming to almost eliminate the problem, and I decided to optimize THIS before doing the actual time alignment. As I changing the Acoustat position, and looking only at the frequency response btw, I simply left the delay compensation as I had previously set it by distance approximation. Starting from the "original" position from a few years ago, which had the 225 Hz suckout, I was getting a step response like this:
Notice the little peak at just over the top of a large hump. That little peak is the beginning of the Acoustat step response, wheras the hump is the first hump of the subwoofer step response.
When I finally decided to move the Acoustats exactly 6 inches further out (for a total of about 39 inches from the wall) because it removed the suckout at 225Hz in the frequency response I was looking at, the step response looked like this:
You can see the little peak of the Acoustat step response moving forwards in time, as the Acoustats are being moved closer to the listening position, just about to the highest point of the first subwoofer hump. But even moved 6 inches forwards (which is about 5 inches closer to the listening position) the subwoofer response still seems to start first. Notice you can also see the little wiggle where the supertweeter is starting about 1 ms later than the Acoustats. At this point I had not yet moved the supertweeters forward to match the Acoustats, which I couldn't completely do anyway, and it appears their initial adjustment was slightly wrong also, but less than 0.5 ms as you can see in the earlier graph where the supertweeter start is a bit more subtle.
And now, for something completely different, here is what the Acoustat step response looks like just by itself (with the low and high pass crossovers I am using):
It pretty much starts with a spike upwards, as it should. Somehow the amplitude appears much greater than it does in combination with the woofer response, partly that's due to normalization and the curious nature of step response computation.
I measured and stared at the subwoofer step response for a long time, and finally made a wild guess about where the beginning of the woofer step should have been computed by REW. I then used the computed panel delay, and MY computed subwoofer delay, to figure out a correct delay compensation. Sadly I was not taking notes about any of this, it was just a guess, and I tried it, and this combined Acoustat and Sub alignment is about good as anything I've done since, which have only been minor excursions away from this point anyway:
The Acoustat step now merges with the woofer step so that the intial rise is a tall spike, as it should be. It still possibly looks like there is some preceding bass response, but that's down by at least 40dB, and is probably just background noise and artifact. I could not make the step look any better by delaying the bass much more than the above. I tried adding a bit more delay, and the step got worse:
OK, so whatever that was, it looks like too much subwoofer delay has pushed the initial Acousat spike down into the hump in some bass noise valley that precedes the actual bass response, which you can see kicking in a few ms after the initial spike with a blast.
Comparing these pictures begins to get subjective again, I admit, and I wish I had something more solid to latch on to. It did seem to me that I could try to maximize the length of the leading edge of the spike. That would indicate that all drive units are contributing maximally to that leading edge. Now, since REW always places the top of the spike at 0dB (btw, I have not yet calibrated the Focusrite interface itself) maximizing the length of the leading edge curiously means that the leading edge should "start" as low as possible. You can see that operating in the above two graphs. In the first graph, what I am calling the noise floor joins the leading spike at -46dB, with the top of the spike at -7dB, so I would count this as showing a leading edge of 39dB. In the second graph, the noise floor joins the leading spike at -30dB, so the leading edge of only 23dB, clearly inferior.
What isn't so easy to interpret, however, is the the negative spike excursion below the joining point means. I think it is mostly artifact, but also should be as low as possible indicating complete cancellation or some such. Anyway, as the delay is adjusted up and down the distance downward changes and also a "gap" seems to open up between the preceding bass-hump noisefloor and the leading edge of the step, making even more a subjective call each time. As I can't remember the actual adjustment for each measurement (my poor note taking on display again) I'm just going to show a bunch of them, in the order I took them, and where I stopped and why.
Trial 3 has leading edge 24dB, very poor.
Trial 4 has leading edge 30dB (but look at how low it goes also--actually to the bottom of the full scale not shown).
Trial 5 has leading edge 28dB, but deep spike again.
Trial 6 has leading edge 47dB and deep valley beforehand, best so far.
Trial 7 has leading edge 32dB, very poor.
Trial 8 has leading edge leading edge 31dB, very poor.
Stepping back in opposite direction now, Trial 9 has leading edge about 35dB.
51dB !!! This appears to be the winner, especially if you take the beginning of the initial spike at -58dB, then the initial spike is 51dB. The supertweeter is already being added in and that might account for some of the wiggle at the start--it appears pretty well aligned too. Alternatively, you could interpret the beginning of the spike as high as -48dB, in which case this is not the winner, but close.
With hand wave you could almost say this is almost looking like a pretty good step response, with a tall initial spike and pretty well filled in right after that.
This is clearly not as good, a step backwards, with the leading edge at no more than 40dB.
After the 11th measurement, I believe I went back to the preceding one, Alignment 10, as the best. I can't be sure of this, because I didn't take notes, so it might have been one of the other pretty good ones, such as The First Guess. But I'm pretty sure it was the 10th, because I could still remember the values for the 10th when I did the 11th, and the 11th was clearly a step backwards.
The curious delays in the DSP boxes to achieve the best alignment are:
Subwoofer: 6.3 ms delay
Supertweeter: 4.6 ms delay
Panels: 5.3 ms delay
So the subwoofers are delayed 1 ms MORE THAN than panels, though they are also about 2 feet further back. That is the second issue I don't understand.
I have already ruled out the lowpass crossover as causing this. It also probably not because of delays in the complex interface box of the Acoustats, because the delays required to align the supertweeters and the panels is about what one would expect based on distance.
This should not be confused with the first issue I don't understand, which is why REW is apparently computing the subwoofer delay time incorrectly. This is an entirely separate issue, as I did my alignment by the graphical means shown above, and NOT by using REW's delay estimates, especially for the subwoofer. However, there might be some connection I cannot yet fathom between the two. Or maybe not.
Now some might find it outlandish to write so much and not say a word about how it sounds. Of course it sounds wonderful! My "electrostatic" bass is now even more well integrated into the sound, without being the least bit less impactful. Bass lines are easier to follow. And in many other ways it sounds better overall.
But you should expect that I would feel this way, and it's probably not best to fully trust the audio judgements of the audio investigator, and probably not the measurements either.
Next month it could be different.
No comments:
Post a Comment