High Resolution Audio: Does It Matter?

[Sal1950]

Precisly this, and this is why I have stuck with stereo rather than multi channel - my favourite recordings are stereo so all the benefits of spatial accuracy of multi channel don't apply to them

Frank you might be surprised if you did some homework on what's available in multich today.
I'm not a classical person but I recently took advantage of a Berliner Philharmoniker free 7 day offer to check out the offerings.
They are streaming some amazing music from live concert events much in some awesome multich sound.

Concerts | Digital Concert Hall

The Digital Concert Hall archive contains over 700 video recordings of Berliner Philharmoniker concerts. The search categories help you find the content of your choice.

www.digitalconcerthall.com

After many years involvement in multich playback I could never be happy limited to stereo only playback.
It's not a matter of money for you, expand your horizons by adding multich capability to your current system.
You won't be disappointed.

 

[j_j]

Quite so, it really doesn't matter what is best or better in theory, in practice my 60 years worth recordings are mostly stereo with quite a bit of early mono so that is what I need a hifi to play.

The 3-channel front is not from theory, it was found first in practice. On the other hand, the marketing flack from the opposition was all centered around the phrase "you only have two ears" ignoring both that they move, and that they sit in a soundfield. It won, of course.

The "center vocal channel" is an artifact of cinema production,which is another problem entirely.

 

[Suffolkhifinut]

Frank you might be surprised if you did some homework on what's available in multich today.
I'm not a classical person but I recently took advantage of a Berliner Philharmoniker free 7 day offer to check out the offerings.
They are streaming some amazing music from live concert events much in some awesome multich sound.

Concerts | Digital Concert Hall

The Digital Concert Hall archive contains over 700 video recordings of Berliner Philharmoniker concerts. The search categories help you find the content of your choice.

www.digitalconcerthall.com

After many years involvement in multich playback I could never be happy limited to stereo only playback.
It's not a matter of money for you, expand your horizons by adding multich capability to your current system.
You won't be disappointed.

Think it was Sony that discredited surround sound many years ago destroying the creditably of SQ / QS four channel music reproduction. Go to a concert and the musicians are ranged in front of the listeners, we don’t sit among them. Probably while I’ve never believed in multichannel music reproduction.

 

[j_j]

Think it was Sony that discredited surround sound many years ago destroying the creditably of SQ / QS four channel music reproduction. Go to a concert and the musicians are ranged in front of the listeners, we don’t sit among them. Probably while I’ve never believed in multichannel music reproduction.

It's much more complicated than that. When moving to multichannel, a panpot is no longer the proper tool.

 

[krabapple]

Oh bless - what is the purpose of our audio systems? I assume we are all music lovers who want to enjoy and appreciate the talents of our preferred musicians.

Or maybe it really is just about the science . Pretty sure the performers on the recordings didn’t have that in mind

Wrong question.

Correct question: what is the purpose of *this website*?

 

[krabapple]

Precisly this, and this is why I have stuck with stereo rather than multi channel

JJ was talking about multi-*mic-ing*, a recording technique that's been in use for decades for stereo classical recordings. Karajan made it (in)famous, IIRC.

(A multimic'ed recording can also be mixed to surround of course. )
(and not all classical recordings are multimic'ed ,of course)

- my favourite recordings are stereo so all the benefits of spatial accuracy of multi channel don't apply to them. As long as I personally am satisfied by instrumental timbre and a dynamic range as realistic as I can get I have gone as far down the road to high fidelity as is available to me.

I count myself among the many (including Dr. Floyd Toole) who get great enjoyment from upmixing stereo recordings to multichannel output via processing that leverages properties intrinsic to the recordings. I'll never go back to 'just' stereo.

 

[krabapple]

Well, the results from Steinburg and Snow in the1930's do show that 3 channels, all in front, are what you want.

But that would require proper 3 channel recordings, which haven't happened since Fantasia.

What about the Mercury Living Presence etc. classics that were recorded in 3 track, mixed down to 2 for years, but eventually released as 3-channel (plus silent channels) SACDs?

(Ditto some jazz classics...e.g. Kind of Blue and In A Silent Way)

 

[j_j]

What about the Mercury Living Presence etc. classics that were recorded in 3 track, mixed down to 2 for years, but eventually released as 3-channel (plus silent channels) SACDs?

(Ditto some jazz classics...e.g. Kind of Blue and In A Silent Way)

They are quite good, adn the real center channel sets up the actual distance cues.

 

[Deleted member 16543]

...
While you may never hear a 21kHz tone, it is indeed possible that a bit of 21 kHz content leaking past the cochlear filter COULD (not "does") affect the detection.
...

I think we can boil down our divergence of opinion on this sentence above.
You say that if a sound wave having content above 20 kHz hits our ears, the 21 kHz content may "leak" (how?) past the cochlear filter, affecting the detection.
I can't understand how, then, you also say that we can't hear tones past 20 kHz (or even 18 kHz, past 40). Isn't the "leakage" a form of audibility?

It seems to me that you believe that, although only in some particular cases, we actually can hear above 20 kHz, because of some ear non-linearities.
If that's the case, I have no problem admitting that higher sample rate is needed.
I just don't see how a "leak" can happen at 21 kHz, which affects the perceived sound, but at the same time a tone at 21 kHz can't possibly be heard.

What is the mechanism that makes the presence of ultrasonic frequencies end up as a potentially detectable sound difference? What kind of signal would showcase this phenomenon most clearly, if a 21 kHz tone isn't it?

 

[Freeway]

On a related but fairly off-topic note, I saw this video a while back and couldn't figure out a place to post it. It's interesting to see what was involved 'back in the day.'

Not just back in the day. Synthesizers @ 2:00 mark.
Thought about posting in 'cables make a difference' discussion.
Anyhow here it is.

 

[j_j]

I think we can boil down our divergence of opinion on this sentence above.
You say that if a sound wave having content above 20 kHz hits our ears, the 21 kHz content may "leak" (how?) past the cochlear filter, affecting the detection.
I can't understand how, then, you also say that we can't hear tones past 20 kHz (or even 18 kHz, past 40). Isn't the "leakage" a form of audibility?

It seems to me that you believe that, although only in some particular cases, we actually can hear above 20 kHz, because of some ear non-linearities.
If that's the case, I have no problem admitting that higher sample rate is needed.
I just don't see how a "leak" can happen at 21 kHz, which affects the perceived sound, but at the same time a tone at 21 kHz can't possibly be heard.

What is the mechanism that makes the presence of ultrasonic frequencies end up as a potentially detectable sound difference? What kind of signal would showcase this phenomenon most clearly, if a 21 kHz tone isn't it?

If you take a tone, and window it, it becomes more broadband, and can have "information" at lower frequencies. This is also related to the talk I'm preparing for a tutorial titled 'what is bandwidth'. We must all remember that short-term clips of sine waves do not consist of a single frequency. This is perhaps part of what you're not understanding (which is to say that when you chop up 21kHz, you can get sidebands well below 20khz if you do it grossly).
The other part is that the cochlear filters are not the same kind of "brickwall" you think of, and they are very importantly active, not passive, and react to stimulii in real time.

For the first part, I'll make a picture in a minute in Matlab.

For the second, watch that video linked above about hearing, and then read my hearing tutorial on the AES PNW site. While the author of the video and I do not agree on exact mechanisms, we do agree on the results. This will help explain the cochlear filter issue.

Now, the first line is a 22kHz sine wave sampled at 192kHz. I used 192kHz simply so you can see what's going on without having to anti-image the plot. Vertical is waveform value, horizontal is time.
The second line is the spectrum. Note it's quite nearly a pure sine wave. Vertical is magnitude of frequency response, horizontal is frequency in kHz.
Third line now shows a short burst of the same frequency, and fourth line shows the actual spectrum of THAT signal. Notice how it's very, very broadband.

So, 22 khz is only 22khz when it's continuous.

Now, consider. In the second case, if it's sent through a brickwall filter, you'll have only the parts below 21khz. All good, but you ARE hearing something (if you have perfect hearing) resulting from a stimulus with a CENTER FREQUENCY of 22kHz.

Now, if you put that ONLY through the cochlear filter, the cochlear filter reacts to the whole signal, not just the brickwall filtered part. So the results are different in waveform, and *** MAY *** be different in perception.

 

[Sal1950]

Thought about posting in 'cables make a difference' discussion.

OMG, all the wires!

 

[Deleted member 16543]

If you take a tone, and window it, it becomes more broadband, and can have "information" at lower frequencies. This is also related to the talk I'm preparing for a tutorial titled 'what is bandwidth'. We must all remember that short-term clips of sine waves do not consist of a single frequency. This is perhaps part of what you're not understanding (which is to say that when you chop up 21kHz, you can get sidebands well below 20khz if you do it grossly).

No no, I'm familiar with the properties of Fourier transform. I have a background in electronics engineering, so the math in not beyond my comprehension. A multiplication (windowing) in the time domain corresponds to a convolution of the spectra of the two signals you're multiplying, that is the spectrum of the original signal and that of the window. This can create sidebands in the audible band.
What I'm not completely convinced of is that this windowing is a complete modeling of the cochlear filters, when we approach the ultrasonic region.
Isn't it at least *possible* (if not in my opinion likely) that, ultrasonic frequencies being inaudible as pure tones as we both agree on, the cochlear filter applies its convolution effect in the frequency domain only to the content below 20 kHz (or 18, or even less, as we age) that hits the eardrum?

The other part is that the cochlear filters are not the same kind of "brickwall" you think of, and they are very importantly active, not passive, and react to stimulii in real time.

For the first part, I'll make a picture in a minute in Matlab.

For the second, watch that video linked above about hearing, and then read my hearing tutorial on the AES PNW site. While the author of the video and I do not agree on exact mechanisms, we do agree on the results. This will help explain the cochlear filter issue.View attachment 215818

Now, the first line is a 22kHz sine wave sampled at 192kHz. I used 192kHz simply so you can see what's going on without having to anti-image the plot. Vertical is waveform value, horizontal is time.
The second line is the spectrum. Note it's quite nearly a pure sine wave. Vertical is magnitude of frequency response, horizontal is frequency in kHz.
Third line now shows a short burst of the same frequency, and fourth line shows the actual spectrum of THAT signal. Notice how it's very, very broadband.

So, 22 khz is only 22khz when it's continuous.

Now, consider. In the second case, if it's sent through a brickwall filter, you'll have only the parts below 21khz. All good, but you ARE hearing something (if you have perfect hearing) resulting from a stimulus with a CENTER FREQUENCY of 22kHz.

Now, if you put that ONLY through the cochlear filter, the cochlear filter reacts to the whole signal, not just the brickwall filtered part. So the results are different in waveform, and *** MAY *** be different in perception.

As above, this makes perfect sense to me in the mathematical modeling. But wouldn't the side-bands below 20 kHz created by the time windowing (I assume that's the behavior of the cochlear filter, on graph 3, you're coarsely modeling) make any signal at 21 kHz, even a continuous tone, audible just as well?
Why do we not hear these sidebands when listening to a 21 kHz pure tone?
This is what makes me suspect that there might be some other mechanism that either stops the transmission of ultrasonic content past the eardrum altogether, or applies the broadening of the band effect only to the content below 20 kHz that gets past it.

Could you post those links to video and tutorial about hearing again, please?

 

[j_j]

This is what makes me suspect that there might be some other mechanism that either stops the transmission of ultrasonic content past the eardrum altogether, or applies the broadening of the band effect only to the content below 20 kHz that gets past it.

Could you post those links to video and tutorial about hearing again, please?

The point is, rather, that there is sub 20khz content in a windowed signal. This isn't a "model'. What I am suggesting that you do is analyze a longish, very sharp FIR filter at 20kHz with (for convenience) a hann window about .5 milliseconds wide, and look at the output of that. Because that's "more or less" (Yes, there are many more details, but this suffices to illustrate) shows what happens.

And, yes, stuff even at 20kHz starts to bounce off the eardrum. That's part of the reason for the rolloff.

As an aside there is near-perfect match between the cochlear filter bandwidths (ERB's) and the information content in speech up to about 8kHz (not that this is surprising, of course), and above that, the system appears to be set up more for direction detection. (again,makes sense), so there is little to be gathered at 20kHz, give the size of our head, pinna, etc.

The hearing tutorial is at www.aes.org/sections/pnw in the "past meeting recaps" area, both powerpoint deck and recording.

 

[Deleted member 16543]

The point is, rather, that there is sub 20khz content in a windowed signal. This isn't a "model'. What I am suggesting that you do is analyze a longish, very sharp FIR filter at 20kHz with (for convenience) a hann window about .5 milliseconds wide, and look at the output of that. Because that's "more or less" (Yes, there are many more details, but this suffices to illustrate) shows what happens.

I'm not sure what you're asking me to do here. Just to make sure we're on the same page..
Do you want me to check what happens to a sinc function (the kernel of a steep, linear phase brickwall filter) when it's convolved with a Hann window type of kernel, in the time domain?
Or do you mean to check the output when the two kernels are multiplied, instead of convolved, shifting the Hann window (one sample at a time, or every .5 milliseconds?) as I examine the various outputs?

And, yes, stuff even at 20kHz starts to bounce off the eardrum. That's part of the reason for the rolloff.

As an aside there is near-perfect match between the cochlear filter bandwidths (ERB's) and the information content in speech up to about 8kHz (not that this is surprising, of course), and above that, the system appears to be set up more for direction detection. (again,makes sense), so there is little to be gathered at 20kHz, give the size of our head, pinna, etc.

The hearing tutorial is at www.aes.org/sections/pnw in the "past meeting recaps" area, both powerpoint deck and recording.

Thank you. Much appreciated.

 

[j_j]

Do you want me to check what happens to a sinc function (the kernel of a steep, linear phase brickwall filter) when it's convolved with a Hann window type of kernel, in the time domain?

I'm asking you to analyze a lowpass filter (a sinc might be worse, though, I'd pick an optimized 20k-22k cutoff remez design at 96k. Keep it to maybe .1dB ripple and 120dB rejection.

But analyze it on a half-millisecond window basis, sample by sample. This gives you a sample by sample response. Look where the half-millisecond window leaves some serious splatter.

Basically make a 2D array of amplitude vs. time, can use Hann windowed FFT to calculate the frequency analysis, and the plot it as a surface, and look around 20kHz, see what you get.

You may be surprised.

This will make the filter in stock matlab:

clc
clear all
close all

fs=96;
fs2=fs/2;
wp=20;
ws=22.05;

len=220;

f=[0 wp ws fs2]/fs2;
a=[1 1 0 0];
bb=firpm(len,f,a, [ .001 100]);
freqz(bb)


which will look like this in the passband:

That is dB ripple of the filter. Divide by 2 to get the usual +- reported results.
Now, next is the entire filter response:

And, then, you analyze that with this code:

alen=fs2;

%cheating here, that makes this .5 millisecond
%because I'm working in khz above

zz=zeros(alen,1);

bb=[zz; bb; zz]; %pads zeros so i can do the analysis in one loop

res=zeros(len,(alen/2+1));
jj=1;
wind=hann(alen);

for ii = (alen+1)len+alen)

t=wind .* bb(iiii+alen-1));

t=abs(fft(t));
res(jj,=t(1alen/2+1));

jj=jj+1

end

clear t
t=max(max(res));

res=res/t;

res=max(res,.001);
res=20*log10(res);

surf(res)

which gives you this surface

left/right is time, front back is arbitrary units of frequency.

Now, you see how the part around the transition frequency spreads out a lot?

THAT is the potential problem. Now the real window is more complex, etc, but this suffices, I hope, to make the point

 

[Frank Dernie]

JJ was talking about multi-*mic-ing*, a recording technique that's been in use for decades for stereo classical recordings. Karajan made it (in)famous, IIRC.

Which ones?
I am only familiar with Mercury and the Decca tree using 3. Most of the others used variations on the spaced omnis and crossed cardioids in my recollection which may be suspect, I am old.

OTOH
I don’t like all the extra speakers and cables needed for multi channel either, any minuscule spatial special effects one may get from applying some sort of algorithm to make multiple channels out of two is faux and not worth owning extra equipment for.

I do have a centre and rear channels for the rare occasion I watch a film so have tried this faux surround but went back to stereo in the end.

 

[Deleted member 16543]

I'm asking you to analyze a lowpass filter (a sinc might be worse, though, I'd pick an optimized 20k-22k cutoff remez design at 96k. Keep it to maybe .1dB ripple and 120dB rejection.

But analyze it on a half-millisecond window basis, sample by sample. This gives you a sample by sample response. Look where the half-millisecond window leaves some serious splatter.

Basically make a 2D array of amplitude vs. time, can use Hann windowed FFT to calculate the frequency analysis, and the plot it as a surface, and look around 20kHz, see what you get.

You may be surprised.

This will make the filter in stock matlab:

clc
clear all
close all

fs=96;
fs2=fs/2;
wp=20;
ws=22.05;

len=220;

f=[0 wp ws fs2]/fs2;
a=[1 1 0 0];
bb=firpm(len,f,a, [ .001 100]);
freqz(bb)


which will look like this in the passband:
View attachment 216001
That is dB ripple of the filter. Divide by 2 to get the usual +- reported results.
Now, next is the entire filter response:

View attachment 216002


And, then, you analyze that with this code:

alen=fs2;

%cheating here, that makes this .5 millisecond
%because I'm working in khz above

zz=zeros(alen,1);

bb=[zz; bb; zz]; %pads zeros so i can do the analysis in one loop

res=zeros(len,(alen/2+1));
jj=1;
wind=hann(alen);

for ii = (alen+1)len+alen)

t=wind .* bb(iiii+alen-1));

t=abs(fft(t));
res(jj,=t(1alen/2+1));

jj=jj+1

end

clear t
t=max(max(res));

res=res/t;

res=max(res,.001);
res=20*log10(res);

surf(res)

which gives you this surface
View attachment 216004

left/right is time, front back is arbitrary units of frequency.

Now, you see how the part around the transition frequency spreads out a lot?

THAT is the potential problem. Now the real window is more complex, etc, but this suffices, I hope, to make the point

Thank you for posting this plot as I, like I suspect many of the people reading this, don't have Matlab.
I had a hunch you meant the second one of my two options (the hardest one to describe, even qualitatively).
In both cases I would have expected some kind of "smearing" of the bandwidth content, which is evident in your plot.
However, your plot also shows a smearing in time, the distribution of which with frequency would have been more difficult (let's say impossible) for me to figure out without a software like Matlab.
I added some notations to your plot, to make sure I'm reading the plot correctly. I do have a couple observations, but I'll wait until I read the material you linked to, because some answers may be there already.

The way I read this plot, time runs right to left (from 0 to 250).
The frequency "smearing" around Fc/2 is indicated in purple, and the one in the time domain is indicated in grey. This time "smearing" is distributed with frequency and mainly concentrated around Fc/2 as well.
I would interpret this, assuming that the mathematical model matches how we actually perceive things throughout the whole audio band, as the ear being able to detect high frequency content before the whole spectrum is perceived, at least under stimuli similar to an impulse.
Basically, at t=0, we perceive a first signal whose frequency spectrum is the first slice of the graph in time, all the way to the right (a wide bell-like shape, centered around Fc/2, highlighted in yellow).
Then, after about 70 time units after, we actually start to perceive the whole frequency content.

Am I reading this right?

 

[j_j]

Thank you for posting this plot as I, like I suspect many of the people reading this, don't have Matlab.
I had a hunch you meant the second one of my two options (the hardest one to describe, even qualitatively).
In both cases I would have expected some kind of "smearing" of the bandwidth content, which is evident in your plot.
However, your plot also shows a smearing in time, the distribution of which with frequency would have been more difficult (let's say impossible) for me to figure out without a software like Matlab.
I added some notations to your plot, to make sure I'm reading the plot correctly. I do have a couple observations, but I'll wait until I read the material you linked to, because some answers may be there already.
View attachment 216056

The way I read this plot, time runs right to left (from 0 to 250).
The frequency "smearing" around Fc/2 is indicated in purple, and the one in the time domain is indicated in grey. This time "smearing" is distributed with frequency and mainly concentrated around Fc/2 as well.
I would interpret this, assuming that the mathematical model matches how we actually perceive things throughout the whole audio band, as the ear being able to detect high frequency content before the whole spectrum is perceived, at least under stimuli similar to an impulse.
Basically, at t=0, we perceive a first signal whose frequency spectrum is the first slice of the graph in time, all the way to the right (a wide bell-like shape, centered around Fc/2, highlighted in yellow).
Then, after about 70 time units after, we actually start to perceive the whole frequency content.

Am I reading this right?

More or less. the Fc is actually the transition band of the filter and the width is fc*2, not fc/2 in frequency, and the width in time is inversely proportional to fc, i.e. the faster the cutoff, the longer the time response. The key to making sure that there is no possibility at all of a problem is pretty simple, move it up a few kHz, and it's all over, so to speak. You can't get detection when it bounced off the eardrum. The mathematical model here is somewhat of a match to something resembling an actual cochlear filter, but this is actually a touch better (in terms of smear), and a few orders of magnitude easier to actually calculate.

The bump at the end also exists at the beginning, but isn't shown here because of when I started and stopped the analysis. It's way, way down.

Later I will show what a filter with a wider transition band is like. But now I'm outside digging up an overgrown patch of horseradish.

 

[j_j]

Ok. I did a filter with the same in band ripple and stop band rejection cutting off between 20k and 32k. Note the scales are a bit different here, the 't' is still in samples like the above but please to note that the total length of the axis is much, much smaller! Instead of zero to 200-some, it's 80.

You will also note that the "spread part" in the middle is wider, and much more similar to the "overall" response.
All in all, a much, much better behaved filter. I also edited the program to plot both sides of the plot completely, so the leading part (near zero) is now complete, unlike the other plot. This is only a plotting issue.