High Resolution Audio: Does It Matter?

[Digital Mastering System]

When I was a young engineer in the early '80s, I sat in on some Zubin Meta - CBS Masterworks recordings and mix downs with the producer Andy Kasdan. Kasdan used all 32 channels available on the recorder and premixed some of those - I think he had 50 mics out there. During mixdown to stereo of the edited multitrack (spliced together from scores of takes), he would ride the gain of the 'soloist' instrument of the time, emphasizing the part. He was also very fond of using a continuously variable high pass filter to get rid of room rumble when there was no low freq music to mask it - I thought that was very effective and improved the recording considerably, despite what you might think of the multiple mic approach. I also sat in on a von Karajan session at the Philharmonie (w/ DG) and he did pretty much the same thing re multiple mic setup.

 

[krabapple]

Of course, multi track recordings but that came much later and,

It was common for DG --- one of the dominant labels then -- by the mid 70s. That's 'much later' than the first orchestral recordings but it is half a century ago now.
Multimicing orchestras remains common, though not in as extreme form as the most extremely multimiked DG's.

Orchestral Recording

REVISITING AND REFINING CLASSIC TECHNIQUES

www.mixonline.com

There are still some purists who fully embrace these proven “old-school” techniques, but it has been much more common during the multitrack age for engineers to augment either the Blumlein or Decca setup (often modifying each) with spot mics over small groups of players or sections of the orchestra that can be brought into the mix later.

(that was in 2006)

of course loses all hall ambience (so fine for studio recordings)

Depends on how it's done. You can add mics for that too.

 

[Deleted member 16543]

For

It was common for DG --- one of the dominant labels then -- by the mid 70s. That's 'much later' than the first orchestral recordings but it is half a century ago now.
Multimicing orchestras remains common, though not in as extreme form as the most extremely multimiked DG's.

Orchestral Recording

REVISITING AND REFINING CLASSIC TECHNIQUES

www.mixonline.com

There are still some purists who fully embrace these proven “old-school” techniques, but it has been much more common during the multitrack age for engineers to augment either the Blumlein or Decca setup (often modifying each) with spot mics over small groups of players or sections of the orchestra that can be brought into the mix later.

(that was in 2006)




Depends on how it's done. You can add mics for that too.

For what it's worth, my preference on this spin off conversation, especially when recording orchestras or acoustic music in a manner that tries to recreate the realism of the performance during playback, is for recordings done with a binaural microphone, played back over pure stereo (hard panned L and R channels), with near field sweet spot, and possibly with some form of cross-talk cancellation.
The 3D effect and realism with such recordings and listening setup are unparalleled, in my opinion.

 

[j_j]

@j_j Is this the tutorial you're talking about?

AES PNW Meeting Report - Auditory Mechanisms for Spatial Hearing

No, you need to go back a few years to something like "Hearing 101" or "Hearing 099" or something like that. You need to start there.

 

[j_j]

I'm not sure what you're referring to here. I see a shorter time "smear" around the cut off frequency, indicated in black.
I also see a more wide band initial detection time, extending all the way down to the bass frequencies, around 10 time units, indicated in green.
However, the bass region looks a lot more rugged after the first detection than in the sharp cut off filter example. It does have an initial rise around 10 time units that's similar to the rise around Fs/2, but then rises and drops a couple times (highlighted in red).
So could we possibly be talking about a trade-off?
I also used the correct english naming convention on the frequency scale.

View attachment 216161

What you're missing is that the "hump" that leads and lags the center is not FS/2. It's the center of the TRANSITION BAND of the filter. The plot is run with a samplign rate of 96kHz, and the filter response is independent of the sampling rate, jsut all below FS/2. Remember, you CAN NOT create anything above fs/2. '11' is the 48k point. 1 is the DC point. There are only 11 frequencies plotted here.

You notice that the response width between the green line and the similar one on the far end is MUCH MUCH MUCH shorter than the one for the steeper filter. This is the whole point. The "roughness" in the bass is not actually roughness, you can ignore it in this case. There's no bass artifact, because the length of the bass frequencies are far, far longer than those peaks and dips that you're seeing sampled by a process that is too short to even resolve bass. So you can ignore that.

The filter starts to roll off at 20khz and finishes rolling off at 32k. The transition band is WIDER, therefore the "spread" is shorter.

This is showing HIGH FREQUENCY effects. Using an appropriate analysis width for a bass frequency would simply be a smooth bump. Only look above about line 3 or 4 for anything meaningful.

 

[j_j]

https://www.aes-media.org/sections/pnw/pnwrecaps/2019/apr2019/ is what you're looking for for the hearing tutorial.

 

[Deleted member 16543]

What you're missing is that the "hump" that leads and lags the center is not FS/2. It's the center of the TRANSITION BAND of the filter. The plot is run with a samplign rate of 96kHz, and the filter response is independent of the sampling rate, jsut all below FS/2. Remember, you CAN NOT create anything above fs/2. '11' is the 48k point. 1 is the DC point. There are only 11 frequencies plotted here.

You notice that the response width between the green line and the similar one on the far end is MUCH MUCH MUCH shorter than the one for the steeper filter. This is the whole point. The "roughness" in the bass is not actually roughness, you can ignore it in this case. There's no bass artifact, because the length of the bass frequencies are far, far longer than those peaks and dips that you're seeing sampled by a process that is too short to even resolve bass. So you can ignore that.

The filter starts to roll off at 20khz and finishes rolling off at 32k. The transition band is WIDER, therefore the "spread" is shorter.

This is showing HIGH FREQUENCY effects. Using an appropriate analysis width for a bass frequency would simply be a smooth bump. Only look above about line 3 or 4 for anything meaningful.

Would it be possible to see both plots with the same time and frequency units?

Showing only 0 to 20 kHz (I believe that would be 1 to 6, in frequency units), would help focusing on the audible band.

For the time axis I would put t0 right before the first deviance of the plot from laying on the t-F plane, if possible.

 

[j_j]

Would it be possible to see both plots with the same time and frequency units?

Showing only 0 to 20 kHz (I believe that would be 1 to 6, in frequency units), would help focusing on the audible band.

For the time axis I would put t0 right before the first deviance of the plot from laying on the t-F plane, if possible.

I guess I can compress the time domain, the rest is a constant anyhow. I should send you a bill


Here, for long filter, required for proper antialiasing at 44.1:

And here, for short filter required for proper anti-aliasing at 64kHz

I have made all the axis match. Horziontal axis is samples. I think you can easily see the change in length with a 48 sample analysis window. Frequency goes zero to48k (that's front to back), and energy goes Bottom (lowest) to top.

In order to avoid excessive length of filter, as well, the filter for 44.1 also shows a bit of a "bump" at the edges in order to reach stop band limits. That part can be removed at the cost of making the filter longer, BUT the point is not that, the point is the width of the main "bump" changing around the transition frequencies of the filter. (note, TRANSITION frequencies, not FS/2 or any such)

Note there is no "bulge" across time for the second filter like you see in the first filter, and remember the lower frequencies are not appropriate to evaluate, because that's the wrong window for low frequency hearing evaluation.

 

[Deleted member 16543]

Also, cutting the frequency domain and removing everything above 20 kHz would help making the plots more meaningful from the auditory point of view.

 

[j_j]

Also, cutting the frequency domain and removing everything above 20 kHz would help making the plots more meaningful from the auditory point of v

Why? That makes no sense.

 

[Deleted member 16543]

Why? That makes no sense.

Just for ease of getting my bearings around the specific frequency values. But I zoomed in and was able to indicate the frequency at which the time "smear" kicks in.
Each slice being 2 kHz, for the sharp filter it is around 16 kHz (assuming we can get rid of the bumps by using a long enough filter), for the gentle filter it is around 26 kHz (8 and 13 slices respectively).
What happens if the sharp filter is actually long enough? As in flatness within +/- 0.005dB, cut off 20 kHz, full rejection (-120 dB) reached at 22.05 Hz. How does the time "smear" curve of the ear model (in green) tend to move? This filter has around 1.5 ms latency (~3 ms long). Does the curve tend to a more straight line throughout a higher and higher frequency, or does it stay more or less the same? In other words, does the point in red, that's now at 16 kHz, tend to move towards 20 kHz?

I find it interesting to note how, with the same scale now being used for both plots, one can see more amplitude variation on the crest in the audio band, as compared to the sharp filter here.

 

[j_j]

Just for ease of getting my bearings around the specific frequency values. But I zoomed in and was able to indicate the frequency at which the time "smear" kicks in.
Each slice being 2 kHz, for the sharp filter it is around 16 kHz (assuming we can get rid of the bumps by using a long enough filter), for the gentle filter it is around 26 kHz (8 and 13 slices respectively).
What happens if the sharp filter is actually long enough? As in flatness within +/- 0.005dB, cut off 20 kHz, full rejection (-120 dB) reached at 22.05 Hz. How does the time "smear" curve of the ear model (in green) tend to move? This filter has around 1.5 ms latency (~3 ms long). Does the curve tend to a more straight line throughout a higher and higher frequency, or does it stay more or less the same? In other words, does the point in red, that's now at 16 kHz, tend to move towards 20 kHz?


View attachment 216301

View attachment 216302


I find it interesting to note how, with the same scale now being used for both plots, one can see more amplitude variation on the crest in the audio band, as compared to the sharp filter here.

The short answer is that the width of the part you indicate is strongly related to the transition bandwidth (wider for narrower transition bandwidth). The part that goes away is the left and right tails.

 

[Deleted member 16543]

I guess I can compress the time domain, the rest is a constant anyhow. I should send you a bill

No kidding! Thank you for indulging me.

 

[j_j]

No kidding! Thank you for indulging me.

For a filter with less passband ripple. As you can see, the spread remains. The filter gets longer and longer.

 

[Deleted member 16543]

The short answer is that the width of the part you indicate is strongly related to the transition bandwidth (wider for narrower transition bandwidth). The part that goes away is the left and right tails.

Right. I was wondering if something like this below happens, where the time "smear" increases (t -->), while the frequency at which the detection instant is the same moves towards 20 kHz. the time axis still runs right to left in this example. The green "t" is just the time "smear". Basically if that hyperbolic looking red profile tends to lay on the F and F=20 axis, as the filter is made more and more accurate (long). Just a curiosity.

The answer seems to be it remains more or less the red curve, no matter the filter length.

 

[Deleted member 16543]

I guess the natural question that comes to mind at this point is what happens if, instead of processing the low pass steep filter with the Hann window, we process the kernel of the corresponding high pass filter (basically, from 0 to 24 kHz you have -120 dB, and from 24 to right up to 48 kHz you have 0 dB)?

 

[j_j]

I guess the natural question that comes to mind at this point is what happens if, instead of processing the low pass steep filter with the Hann window, we process the kernel of the corresponding high pass filter (basically, from 0 to 24 kHz you have -120 dB, and from 24 to right up to 48 kHz you have 0 dB)?

You'd probably see some leakage, depending on the filter. 24kHz really has no part in this discussion, except as fs/4, it doesn't really matter except as an anchor point to locate the center of the spectrum.

More appropriate would be to use an actual high frequency cochlear filter, which I have done, but which I can't reveal thanks to the fact it's quite proprietary information.

The results, however, would be slightly worse, in terms of 'a bit more excitation before, and even more after'. That's all I can really say. So I am understating the effect slightly.

 

[j_j]

Right. I was wondering if something like this below happens, where the time "smear" increases (t -->), while the frequency at which the detection instant is the same moves towards 20 kHz. the time axis still runs right to left in this example. The green "t" is just the time "smear". Basically if that hyperbolic looking red profile tends to lay on the F and F=20 axis, as the filter is made more and more accurate (long). Just a curiosity.

View attachment 216311

The answer seems to be it remains more or less the red curve, no matter the filter length.

That's because what it's related to is the transition bandwidth of the filter, or more strictly the inverse thereof. Even more strictly, 1/passbandwidth + 1/transition band +1/stopbandwidth in which this case the transition band will dominate pretty much entirely. That's not a precise summary, either, but it's a decent rule of thumb. Its also very likely that the width will be in the same frequencies. You rarely see anything in a stop band because the resulting energy is so small.

 

[j_j]

Getting back to my point, which is that we should have just started with an Fs that "makes it go the (*(&* away, thank you".

 

[anmpr1]

When I was a young engineer in the early '80s, I sat in on some Zubin Meta... During mixdown to stereo of the edited multitrack (spliced together from scores of takes), ...

YT has the Golden Ring video chronicling the making of the Solti's Ring. I went back and reviewed it, as I'd forgotten specifics. What interested me, and what I was surprised to find, was how edits/splicing appears to have been done at 7ips, and not 15. One deck has 2 inch wide multichannel tape, but the working tape that they supposedly cut the record from was quarter inch... at 7ips? Sure looks that way. It's not clear from the narrative the detailed steps that were used, however watching the video shows the splicing technique.

I'm amazed they could make such precise splices manually, but I guess with practice it becomes fairly routine. At about 39:00 you can watch one tape being cut, inserted into the splicing block, and joined together on to a new reel.

It is worthwhile to compare Solti with the Keilberth 1955 Bayreuth performance. Evidently using six microphones (three in the pit and three hung over the stage on a lighting scaffold). Political scuttlebutt is that Culshaw deep sixed the Keilberth stereo recording secondary to his own opportunistic purposes. I wasn't there for the negotiations, so I'm only reporting what others have claimed.