Is my mind actually deceiving me?

X

xonik

[H]F Junkie
Joined
Jun 20, 2001
Messages
10,318
After all these years of working with soundcards good and bad, I've finally accepted analytically and empirically the advantage to upsampled content beyond 44.1 kHz, particularly integer multiples like 88.2 kHz, 176.4 kHz, etc.

Now I have been introduced to the Nyquist Theorem and once again, I feel my naivete coming out again.

To those who aren't familiar with it, the Nyquist theroem states that the sampling frequency should be more than twice the frequency of the analog input signal. Scientists have measured the human perception of sound to range from 5 Hz to 22 kHz, as an optimistic figure. Okay, so double 22 kHz and the sampling frequency should be more than 44 kHz, then. 44.1 kHz sounds nice, but wait, that's the sampling frequency of a Redbook audio CD. I think we're on to something...

Now just where does this upsampling fit in? At a resampled 88.2 kHz, by the Nyquist Theorem, we should be able to perceive 44 kHz sounds, which of course can't be the case...right?

Does anyone have any links to research that explains why upsampled content just sounds better, yet still complies with the ideas of the Nyquist Theorem? I've passed A-B tests with upsampled content vs. the original content, but is that good enough?
 
I know very little regarding this Nyquist Theorum, although i found your post interesting. Simply sampling a sound at double its input does not allow you to hear twice the frequency. The human ear can hear typically up to 22khz as more of a physical limitation of the ear drum, and / or that really little bone in your ear which vibrates in accordance with the sound you hear.

I know I don't honestly grok what i am talking about, but let me put it to you another way, a way i seem to be better at - analogy.

When standing by your dog, a whistle is blown. This whistle happens to be a dog whistle, which uses a much higher frequency than a human can hear to call the dog. While the dog can hear it, human cannot. The input frequency into human and canine ear is the same, but human ear has a lower response limitation than that of man's best friend.

Again, if i am totally off on this, see my second paragraph. It's early morning for me.
 
You are overlooking the fact that Nyquist has done the damage to CD audio already (55kHz is the minimum sampling rate necessary to encode all audible sounds) and you can't get back what is not there so when you resample an audio stream that does not have ultrasonic frequency data present what are you changing?







Besides the effectiveness of interpolation? ;)
 
Where did you get the 55 kHz figure? I was under the impression that

f_sample > 2 * f_signal

and that's all there is to it.

And is there any support from the scientific world for the effectiveness of interpolation at frequencies well beyond the detection range of the ear?
 
How about you answer my question before you start asking a bunch of your own? :D

when you resample an audio stream that does not have ultrasonic frequency data present what are you changing?



I did have a better reply typed up but clicking an AIM link kinda killed that and I was to lazy to retype it.
I will get to your questions later. ;)
 
This must be a trick question, because under the Nyquist theorem, any sampling frequency beyond double the signal frequency results in an exact reconstruction of the frequency when converted back to continuous time.
 
Nope, no trick questions.
Redbook audio has already been converted to a digital format so upsampling only changes the way the information in the file is processed...nothing else... you don't change the data and add something that is not present so lets just forget all about Nyquist here for a sec and concentrate on how or why up-sampled audio sounds better.


To answer the other question we have to assume your information is wrong.
As far as the perceived range of human hearing goes.... Recent studies have demonstrated that we do in fact perceive audio will into the 26k range.
AES preprint 3207 showed us that reproduced sound above 26 kHz induces activation of alpha electroencephalogram rhythms and suggested that these can affect perception of sound quality. If we agree with their findiungs or not is moot because a small group of manufacturers known as the DVD forum bought into it and this is part of the reason behind the 55k thing.
If you undo the effects of compression and use Nyquist to calculate the EFFECTIVE sample rate of a DVD audio stream what do you come up with?
I come up with 55k but I overlooked a lot of stuff when I calculated it so.....
 
I'm an Electrical Engineer... perhaps I can help. ;)
------------------------------------------------------------------------------------------------

When a signal is sampled at the Nyquist rate you cannot recover the original signal back because an ideal low-pass filter does not exist (The low-pass filter is used to get the original signal back from the samples). A practical solution is to sample the signal at a rate higher than the Nyquist rate as this allows for a low-pass filter with a gradual cutoff. Still, this can NEVER be perfect, because low-pass filter gains still aren't exactly zero beyond the desired values.

Still more problems exist from aliasing. The Nyquist Theorem is for band-limited signals, which all practical signals are time-limited. A time-limited signal is not a band-limited signal, and vice-versa. All time-limited signals have an infinite bandwidth which messes up the spectrum by overlapping cycles. Of course, to combat this, we use a antialiasing filter, which is still not ideal, so aliasing still occurs.

In summary, practically you CANNOT perfectly recontruct a signal from samples. The higher the sampling rate is, the better your recontruction will be.

I tried to make this as simple as possible. The main point here is that the Nyquist Sampling Theorem is assuming everything is "ideal" or perfect, which is certainly is not.

As a side note: Human hearing is approximately 20Hz through 20kHz. This varies between sex (females hearing higher frequencies) and age. As we age, we lose the higher frequencies.


I've done a lot of design in audio amplifiers... sadly, I'm only now starting to use my designs for myself. :eek:
 
Okay, but what, then, is the technical explanation for why interpolation of upsampled source content just sounds better? 96 and 192 kHz sample rates are well beyond the abilities of humans when speaking from a straight frequency comparison. What, then, is it that makes interpolation at these frequencies effective? I don't need to be convinced that upsampled content is better, but rather, why it makes sense from an EE/audio perspective.
 
K, I know just about nil about audio, other than I like it when it sounds good. ;)

Just from what I've read 'somewhere' (i really have no idea, my brain is a sponge), I was under the impression that higher sampling alows filters to be more gradual and 'transparent', which was already mentioned.

But, I also read something about the slope of the waveform being being more accurately reproduced as the frequency increases, as well as some speculation about ultrasonic frequencies having some effect on the listening experience apart from any distinguishable tones.

But, that could easily be a load of crap, too. :rolleyes:

Just throwing what I've heard out there.
 
I think you are referring to harmonic overtones?
I don't really know what everything here means, but I'm tuning here to see if I can learn more :D
 
xonik said:
Okay, but what, then, is the technical explanation for why interpolation of upsampled source content just sounds better? 96 and 192 kHz sample rates are well beyond the abilities of humans when speaking from a straight frequency comparison. What, then, is it that makes interpolation at these frequencies effective? I don't need to be convinced that upsampled content is better, but rather, why it makes sense from an EE/audio perspective.
Click to expand...


Those 96kHz and 192kHz are sampling rates... not the frequencies you are going to hear. Those are the frequency at which the analog signal is sampled at, which should be more than twice the highest analog frequency that we can hear.

Think of it this way...

---------.----.---------------------.
------.----------.------------- .-------.
---.---------------.----------.-----------.
-.-------------------.------.--------------.
------------------------.-------------------.


Picture those dots as an audio signal. Now, if you could connect the dots, you'd see what the original analog signal looks like. The sampling frequency means how many samples of the analog signal are taken per second. The dots are the samples of the analog signal, which are quite higher than the frequency of the analog signal.

Just for side note: A unsampled source is not interpolated. Interpolation means that you are recovering a signal back from its samples.

As to why higher frequencies sound better, refer to my previous post. The reduction of aliasing and non-ideal low-pass filtering at higher sampling frequencies are the reasons why higher sampling rates sound better.
 
How about an analogy? You could say real sound is sampled as if it was an infinite rate. Nyquist allows for double the frequency just so that all of the frequencies in the range of 20-20000Hz are accounted for. But the detail of the sound is still "blocky" compared to what it would sound like if it was real. Even if all you could hear is the 20-20000 range, and all the frequencies are accounted for, the sampling rate also determines the detail of the sound. Taking samples at intervals of 1 second would sound like beeps, and intervals of billionths of a second would sound so sharp and detailed, it might sound real. Picture this: Most people are not able to tell the difference between displays at 200fps and 60fps. Does this mean that they can't tell the difference between 200fps and infinite fps (real life)? Of course they can! No one is going to mistake looking into a monitor for real life. I know there are limitations in the output devices themselves that will limit realism, but there are benefits to going beyond what is supposed to be the "limit" for what people are supposed to be able to tell.
 
Synful Serenity said:
Nyquist allows for double the frequency just so that all of the frequencies in the range of 20-20000Hz.
Click to expand...

The Nyquist Sampling Theorem was not "invented" for human hearing; it's for reconstruction of ANY analog signal, no matter what the frequency happens to be. All it says is that for a signal to be reconstructed without error, even if that signal is 1Ghz or 1Thz, you need to use a sampling rate of more than twice the signal frequency. The theorem of course is only for an ideal world.


And taking samples at 1 per second would not sound like beeps. It would sound more like a tick.. tick... tick... And that's assuming you could hear it due to the small finite duration of the sample, which would be of a higher frequency than humans could hear. A beep would be a tone signal, which would be a pure sine wave.
 
sEErZer0, I understand everything you're saying except band- and time-limiting, but you still don't explain how the ear can recognize the difference in sampling rates, especially because the differences are somewhat blurred by the interpolation at the DAC stage.
 
Synful Serenity said:
Picture this: Most people are not able to tell the difference between displays at 200fps and 60fps. Does this mean that they can't tell the difference between 200fps and infinite fps (real life)? Of course they can! No one is going to mistake looking into a monitor for real life. I know there are limitations in the output devices themselves that will limit realism, but there are benefits to going beyond what is supposed to be the "limit" for what people are supposed to be able to tell.
Click to expand...
That's just an analogy. I know what my ears are telling me, you don't have to convince me that the upsampled sound is more natural sounding. But you can't make that quantum leap and assume that the hearing system is the same as vision, without support for that claim. Maybe you're right, but is it fair to assume that?
 
Higher sampling rates sound better because they make a better representation of the original analog signal. Differences aren't blurred... they add up. By the time you hear the music, it has gone through many non-ideal stages that add to the distortion. Sampling is only an approximation of the original signal. The ear can recognize it, because the sound... basically sounds like $hit. From the overhanging spectral tails to the non-ideal low-pass filters, it changes the sound of certain frequencies in the music. That along with amplifiers not having equal amplification across the spectrum messes up your music. All these defects add up.

I suggest if you want to hear what I'm talking about, learn Matlab, and start sampling music at higher and lower rates... of the same music. You'll hear the difference. You'll hear the clarity and vibrance of the higher sampling rate. Either that, or download some "legal" music that was sampled at a higher and lower rate, and listen to them.
 
So it sounds like there isn't much of an objective explanation in this area. I know MATLAB, so maybe I'll give that a shot.
 
xonik said:
So it sounds like there isn't much of an objective explanation in this area. I know MATLAB, so maybe I'll give that a shot.
Click to expand...
*Higher sampling rates sound better because they make a better representation of the original analog signal.*

That's reality.
 
Are you in school for engineering or something? Is that where you were introduced to the Nyquist Sampling Theorem?
 
r00k said:
...The human ear can hear typically up to 22khz as more of a physical limitation of the ear drum...
Click to expand...
Its been proven that the human ear can hear frequencies up to at least 200Ghz if its loud enough, according to the guineuss book of world records.
 
Just for side note: A unsampled source is not interpolated.
Click to expand...

How do you figure?
We must be able to interpolate the signal between samples or upsampling is pointless ;)


Edit: now my reply makes sense. :rolleyes:
 
Mister X said:
How do you figure?
We must be able to interpolate the signal between samples or upsampling is pointless ;)


Edit: now my reply makes sense. :rolleyes:
Click to expand...

Read my statement again... you don't need to interpolate an unsampled source. If it's not sampled.... what are you interpolating?
 
lol.
WTF was I thinking?
See what happens when you try to solder and type at the same time? :D
 
Mister X said:
lol.
WTF was I thinking?
See what happens when you try to solder and type at the same time? :D
Click to expand...

You're just lucky you didn't burn your finger off! Hope you didn't buy a Cold Heat soldering iron, that thing sucks. Mine broke within two weeks.
 
Na, no gimicks like that for me.. unless you count my Metcal SP200 as a gimick.
 
You must log in or register to reply here.
Back
Top