There's Life Above 20 Kilohertz!

[jayfrigo]

Originally Posted by 12ax7

With all due respect to everyone (and their arguments) here, the fact remains that there are folks who -- though they can't hear these frequencies -- can tell the difference when these frequencies have been eliminated and/or altered.

Like the Geoff Emerick/Neve story: I don't know what Geoff was hearing, but he heard something, and correctly identified the channels it was coming from.

...And after Rupert found and fixed it, he was happy.

Good anecdote about the improper termination on two channels, but what he was hearing can be traced back to distortion by-products in the audible range.

Also, the science behind digital audio, both theoretical and practical, is well-understood and tested in engineering circles. The same old misunderstandings or incomplete pictures are presented time and time again on various internet forums. It may seem counter-intuitive at first, but if you understand the complete system, and how it really works, you'll see that it can do just fine within the well-known limitations, bandwidth certainly being one of them, and filter issues being another. Many problems so often attributed to it are simply not accurate.

[kjg]

seeing this thread came up again, this song title popped up in my head..

YouTube - Frank Zappa - The Torture Never Stops - HQ

[Audiop]

Originally Posted by UnderTow

Unfortunately it is wrong.

Alistair

Even very wrong!


/Peter

[syncussion]

Originally Posted by UnderTow

Unfortunately it is wrong.

Alistair

Originally Posted by Audiop

Even very wrong!


/Peter

No I don't beleive I'm wrong here.
My explenation was perhaps a little hard to follow reading it back now, but it is right.
What part of it do you think is wrong?
Please elaborate instead of just saying I'm wrong.

[Jon Hodgson]

Originally Posted by syncussion

No I don't beleive I'm wrong here.
My explenation was perhaps a little hard to follow reading it back now, but it is right.
What part of it do you think is wrong?
Please elaborate instead of just saying I'm wrong.

Ok, well the first part is wrong quite simply because you're picking an example that falls OUTSIDE of the shannon-nyquist requirements.

It's a common mistake, using a signal at half the sampling rate in order to attempt to show a deficiency in Nyquist, but actually it just shows a deficiency in your understanding of Nyquist... the sample rate must be MORE than two times the frequency of the highest component, only infinitessimally so in theory (in practice a little more due to filter issues), but always more, so any argument based on a 10kHz signal and a 20kHz sample rate is moot before it starts, since the signal does not meet the constraints set by Nyquist.

The second part is just plain wrong, the ADC captures nothing of a 999Hz waveform until one second?... erm, nope, even allowing for you wording it badly I can't see how you're anywhere near correct.

[Audiop]

Originally Posted by syncussion

No I don't beleive I'm wrong here.
My explenation was perhaps a little hard to follow reading it back now, but it is right.
What part of it do you think is wrong?
Please elaborate instead of just saying I'm wrong.

Well nothing in the post makes sense. No offense. If i find time later I'll write more, but I'm sure someone else with knowledge will go thru it.


/Peter

[syncussion]

Originally Posted by Jon Hodgson

Ok, well the first part is wrong quite simply because you're picking an example that falls OUTSIDE of the shannon-nyquist requirements.

It's a common mistake, using a signal at half the sampling rate in order to attempt to show a deficiency in Nyquist, but actually it just shows a deficiency in your understanding of Nyquist... the sample rate must be MORE than two times the frequency of the highest component, only infinitessimally so in theory (in practice a little more due to filter issues), but always more, so any argument based on a 10kHz signal and a 20kHz sample rate is moot before it starts, since the signal does not meet the constraints set by Nyquist.

The second part is just plain wrong, the ADC captures nothing of a 999Hz waveform until one second?... erm, nope, even allowing for you wording it badly I can't see how you're anywhere near correct.

Aah ok.
I see my messy explenation is the reason for misunderstanding here.
Yes you're absolutely right. But it is not how I ment it.
Indeed if you sample a signal at twice it's rate, 10khz sampling at 20khz than this does not meet Nyquist constraints. I know this and I ment it only as an example to ease understanding of what happens when you sample at slightly higher than twice the frequency.
For instance sample a 9999hz signal at 20khz, which does meet Nyquist constraints. Now you will have a sinal volume envelope on the sampled signal that if you look at the pure digital data will read 0 amplitude to full amplitude every second.
You get a volume envelope to some degree and with varying speeds for all sampled frequencies, even if you sample 20hz at 20khz, though the amplitude envelope will be extremely minimal and easy to reconstruct the original sine with so minimal time smearing to not be important. But for all high frequencies (frequencies nearer half the sampling rate) major time smearing is needed to remove this volume envlope.
I did not mean you get a second of silence every second when sampling 9999hz at 20khz. I ment the adc converter samples a split moment of silence every second and a moment of full scale every second.


edit: btw I forgot to mention in my previous post that the reason non oversampling dacs are bad is not because of the ultrasonic images they give, but because they do not correct these volume envelopes resulting from the phase of the sampled frequency to the sampling frequency.
So non oversampling dacs are bad because they do not smear time, yet they are great in the time domain because they do not smear time
Though I'll take an good oversampling dac over a non oversampling dac any day, it would be great if dacs were made that do both right.

[Jon Hodgson]

Originally Posted by syncussion

Aah ok.
I see my messy explenation is the reason for misunderstanding here.
Yes you're absolutely right. But it is not how I ment it.
Indeed if you sample a signal at twice it's rate, 10khz sampling at 20khz than this does not meet Nyquist constraints. I know this and I ment it only as an example to ease understanding of what happens when you sample at slightly higher than twice the frequency.
For instance sample a 9999hz signal at 20khz, which does meet Nyquist constraints. Now you will have a sinal volume envelope on the sampled signal that if you look at the pure digital data will read 0 amplitude to full amplitude every second.
You get a volume envelope to some degree and with varying speeds for all sampled frequencies, even if you sample 20hz at 20khz, though the amplitude envelope will be extremely minimal and easy to reconstruct the original sine with so minimal time smearing to not be important. But for all high frequencies (frequencies nearer half the sampling rate) major time smearing is needed to remove this volume envlope.
I did not mean you get a second of silence every second when sampling 9999hz at 20khz. I ment the adc converter samples a split moment of silence every second and a moment of full scale every second.


edit: btw I forgot to mention in my previous post that the reason non oversampling dacs are bad is not because of the ultrasonic images they give, but because they do not correct these volume envelopes resulting from the phase of the sampled frequency to the sampling frequency.
So non oversampling dacs are bad because they do not smear time, yet they are great in the time domain because they do not smear time
Though I'll take an good oversampling dac over a non oversampling dac any day, it would be great if dacs were made that do both right.

You're making the second common mistake... looking at the sampled signal and not understanding what makes it up.

The apparant amplitude modulation of the 9999Hz signal sampled at 20kHz when looked at visually is nothing of the sort, it is what you get when you add together a 9999Hz Signal, a 20001Hz Signal, a 39999Hz Signal, a 40001Hz Signal and so on, ad infinitum.

That is the nature of the Z domain, it maps to the continous time with a repeating frequency spectrum, and is why we have anti-aliasing and reconstruction filters. The original 9999Hz component is sampled intact, and preserved as such, but when you look at the sampled signal you see it and its various mirrors superimposed, which confuses the issue... filter out those mirrored components (as in a reconstruction filter) and what is left is your orginal signal, no smearing has occured.

I'm afraid you need to go back a few steps in your understanding and reasoning, because you've misunderstood and are going down the wrong path... and worse still from my point of view, you're propagating your errors to others.

[Audiop]

Originally Posted by syncussion

But for all high frequencies (frequencies nearer half the sampling rate) major time smearing is needed to remove this volume envlope.

A typicall AD/DA link is phase linear and what you say is not correct. If you use a minimum-phase filter then you will get some in band phase distortion though.


/Peter

[syncussion]

Originally Posted by karyoky****

but i would love to know what is meant by "1 hz volume curve" and "1 khz volume curve".

Sorry that should have been 9999 hz (instead of 999hz) sampled at 20khz will give a 1 hz volume curve.
Sorry for the terminology, I'm sure there's a better word for it than volume curve but I don't know it
But what I mean is that the phase of the sampled frequency and the sampling frequency are sort of "beating" against eachother.
Every second the phase of the for instance +-1volt sampled 9999hz frequency will be at a point where just about only 0 volt is seen at the exact time it is sampled, and every second the phase of the +-1volt sampled 9999hz frequency will be at a point where just about the full +-1volt part of the sampled wave is at the exact sampling timing of the 20khz sampling frequency. The rest of the second in time the voltage beeing sampled is between this +-1volt and 0volt.
I know the actual technical implementation of an ADC converter (and DAC) is different but it's just a different form of the same underlying principle so this example still works in showing the information defeciency of sampling at slightly more than twice the frequency, to reconstruct both the frequency, volume and time domains perfectly.

[syncussion]

Originally Posted by Audiop

A typicall AD/DA link is phase linear and what you say is not correct. If you use a minimum-phase filter then you will get some in band phase distortion though.


/Peter

Yes phase linear that's true.
I don't mean changing the phase.
What a typical AD/DA link does is far worse.
I pre rings and post rings. This is the way time is smeared.

edit: this is needed to remove the "volume curves". For continues sine waves the slight pre ringing and post rining when a sine starts and stops doesn't realy matter. But for short percussive / transient sounds it does matter a lot. the smearing gives extra energy to high frequencie transients in a way and many other strange things.
It's just a defecient design overall in my opinion that can never reach practical perfection.

[Audiop]

@syncussion

As Jon said you are on the wrong path. You need to read up on the subject becasue you are plain wrong.

Yes phase linear that's true.
I don't mean changing the phase.
What a typical AD/DA link does is far worse.
I pre rings and post rings. This is the way time is smeared.

But it can't smear time and be phase linear in the passband. All frequencies in the passband comes out with the right relation in phase and amplitude (in a well designed AD/DA link). Yes, it pree ring but at a frequency typically not heard and not on typicall sounds. You need significant transient enery high up in the passband in order to see any ringing and just because something can be seen on a scope doesn't mean it is audible.

edit: this is needed to remove the "volume curves". For continues sine waves the slight pre ringing and post rining when a sine starts and stops doesn't realy matter.
But for short percussive / transient sounds it does matter a lot.

Yes, but only if containing significant energy around the nyquist frequency.

the smearing gives extra energy to high frequencie transients in a way and many other strange things.

No, the energy is not increased in the upper range frequencies.

It's just a defecient design overall in my opinion that can never reach practical perfection.

Most people can not hear the effects on a signal from passing a HQ AD/DA link.. so obviously you are wrong and base your reasoning on a flawed idea of how things are and also from what your eyes see on a scope.


/Peter

[syncussion]

Originally Posted by Audiop

@syncussion

As Jon said you are on the wrong path. You need to read up on the subject becasue you are plain wrong.

I think it may be the other way around.
I'm right and maybe you need to read more

Originally Posted by Audiop

But it can't smear time and be phase linear in the passband. All frequencies in the passband comes out with the right relation in phase and amplitude (in a well designed AD/DA link). Yes, it pree ring but at a frequency typically not heard and not on typicall sounds. You need significant transient enery high up in the passband in order to see any ringing and just because something can be seen on a scope doesn't mean it is audible.

Yes you can smear time and be phase linear.
It's exactly what the pre and post ringing does.

And the ringing frequency is not a fixed frequency or an ultrasonic frequency.
And it's not a design error that it rings or anything.
It needs to ring in order to reconstruct the frequency / volume domain.
And it rings with the frequencies of the signal beeing reconstructed.
This ringing is very very very audible, and it needs to be. If it didn't ring it would be a non oversampling dac!
Now these non oversampling dacs sound very different, hence ringing is audible.
If the ringing wasn't audible what would even be the point of all this oversampling and digital filters etc etc.
The only effect otherwise would be to remove the ultrasonic images that result from not filtering.

Now ringing is a good thing in many cases.
But it is not a good thing in many other cases, for instance very short transient sounds.
Yes this is audible, yes this is measurable, no they don't put this in the specs.

Originally Posted by Audiop

Yes, but only if containing significant energy around the nyquist frequency.

Yes the effect is strongest close to the Nyquist frequency, but it's still very much there at half the Nyquist frequency.
And music has a lot of high frequencies, it's very relevant.
Sure you can say the overall effect of the errors is minor, but where minor improvements are the difference between a $100 dac and a $1000 dac or a $5000 dac it's still very relevant and audible on the right amp and speakers.

Originally Posted by Audiop

No, the energy is not increased in the upper range frequencies.

Yes it often is depending on the sound and what it's phase to the sampling frequency happens to be, especially for short transient sounds, simply because they last longer (due to pre and post ringing).

Originally Posted by Audiop

Most people can not hear the effects on a signal from passing a HQ AD/DA link.. so obviously you are wrong and base your reasoning on a flawed idea of how things are and also from what your eyes see on a scope.


/Peter

Well with the right amp and speakers I can hear the effects.
Many people can't hear the difference between a 128 or 192 kbs mp3 and a wav either.
Not a good argument.
16bit 44.1khz adc dac is flawed and not because of the 22khz frequency limit.
Just as mp3 is flawed in a different way.

[Jon Hodgson]

Originally Posted by syncussion

I think it may be the other way around.
I'm right and maybe you need to read more

That just makes you wrong twice.

[syncussion]

Anyhow.
How this all relates to this thread.

Ultrasonic frequencies (above 30khz) are NOT important.
Infact they can only be bad.

Sampling frequency very high (for instance 192khz or 384khz) is GOOD, if used to sample and reproduce frequencies up till 30khz only with (not like current ad/da works).
This comes at the costs of extra processing power needed for instance to remove DC, but so be it.

If there wasn't so much confusion and marketing hype we'de all be having cheaper and much better sounding converters.

[Plush]

Originally Posted by Kees de Visser

Plush, do you roll off before or after the ADC ? Do you also LP filter 96kHz recordings ?
I can imagine that reducing AD/DA artifacts (aliasing/imaging) at 44.1 kHz rate might be benificial but at 96 and higher rates they "shouldn't" be audible.
If these artifacts are relevant, your filtered version could actually sound closer to the analog source, so comparing to the unfiltered digital version is perhaps not fair.

Hello Kees,

I am just talking about 44.1 KHz work. That is where I do the roll-off. At 96 I just leave it flat of course. I do the roll off before the a/d with an original (1988) Focusrite ISA 110.

For the other posters asking why I can hear that it is more pleasant, I posit that it is not related to frequency response but rather a lack of aliasing noises.

[syncussion]

Originally Posted by karyoky****

hi,

i do not know where you are going with the "volume curve" terminology.

but anyhow......... assuming [erroneously, of course] that the sampler has zero jitter, zero aperture error, and that the sample intervals are absolutely perfectly spaced, then there is only one frequency that will contain the instantaneous amplitude values that are obtained when a signal is sampled [if the signal complies with the requirements of sampling theorum]. ideally, by the terms of the theorum, you do not have to sample each continuous point of the wave's [or wavelet's] rise and fall around the zero point to be able to determine its frequency.

Yes you're right.
Even if you're sampling a +-1volt sine with a frequency of 10khz with a sampling frequency of 20,000.01 hertz you can reconstruct the original 10khz sine perfectly because only the 10khz sine will hit the sampled points exactly.
BUT only if you have this exact sine perfectly for 100 seconds. Here time comes into play.
If you sample the 10khz sine for only 10 seconds at 20,000.01 hrtz there is no way possible to reconstruct the original sine accurately again.
Btw I'm calling this reconstruction "ringing" as it's often described. Again ringing is not an error it's an essential part of the process.
Now if you reconstuct the sampled data in this way then you get problems with fast high frequency transients. There simply isn't enough data to reconstruct them.
It's a long long story to fully explain and i'd have to post graphs etc.

Originally Posted by karyoky****

i believe that analog to digital converters generally use a "sample and hold" method of acquiring the signal amplitudes.

there are obviously defects in the process, and the sample intervals are not going to be perfectly evenly spaced, and so forth, yadda, yadda, yadda.

Yes but no matter how it's implemented it retains the problem that there's not enough information to reconstruct everything perfectly, even besides the errors (which are greatly exadurated in the "reconstruction process")

Originally Posted by karyoky****

and yes, there is ringing other than at the nyquist frequency. ringing can be caused by a number of things. anything that involves filters involves ringing. filters are not perfect. apparently there is a lot of interest among designers lately in developing and implementing techniques that minimize ringing.

dcs converters have selectable filters, where you can choose more or less ringing.

As I said before, ringing is not a side effect or an error but it is an essential part of the reconstruction of the original wave. Less ringing brings less time smearing but worse frequency/volume accuracy.

Originally Posted by karyoky****

amplitudes are not accurately captured by samplers [quantizers]. so you get quantization error.

here is a link that may be of interest:


Analog Sampling Basics - Developer Zone - National Instruments

for what its worth, i notice that they recommend a bandwidth of 3 to 5 times the highest signal of interest, and a sample rate of 5f in order to "more accurately reproduce the waveform".

Thanks for the link.
And it's funny that they recomend a higher sample rate to more accurately reproduce the waveform, like I did a few posts ago
It will indeed more accurately reproduce especially the transients, compared to modern oversampling ad/da.
And I maybe from an engineering point I must agree with bandwidth 3 to 5 times the highest signal of interest, but only because of analogue filter design considerations of the ADC.
So this would bump up the total sampling frequency to 30khz x4 x5 = 600khz

[Jon Hodgson]

Originally Posted by karyoky****

for what its worth, i notice that they recommend a bandwidth of 3 to 5 times the highest signal of interest, and a sample rate of 5f in order to "more accurately reproduce the waveform".

Yes, but that's based on a misrepresentation that your resultant signal consists of the sample points connected by straight lines, and they're not necessarily including anti-aliasing filters.

This makes more sense if you consider that they are talking largely about using sampling for measurement, not audio reproduction.

All in all it's not a particularly well presented document, and is very easy to misunderstand.

[kjg]

Originally Posted by syncussion

So this would bump up the total sampling frequency to 30khz x4 x5 = 600khz

you are off by about a factor 10.

20kHz is really about the upper limit of hearing, and two samples per period is really enough (Nyquist *was* right). 60kHz means a Nyquist frequency of 30kHz which allows half an octave for AA filtering (20-30kHz), which is enough to have a smooth filter.

try reading this paper from Dan Lavry:
http://lavryengineering.com/forum_im...ing_Theory.pdf

[Jon Hodgson]

Originally Posted by karyoky****

hi,

no, i think the "straight line" graphs are depictions of the outcome of sample rates less than what are required by the sampling theorum. maybe i missed the part you are referring to?

"if you increase the sampling rate to 2f, the digitized waveform has the correct frequency (same number of cycles) but appears as a triangle waveform"

Of course they're using the dreaded signal at nyquist there, already a no-no when trying to explain things, they also later refer to anti-aliasing filters as optional, which they may be when you're dealing with measurements of slow signals, but they are not when dealing with digital audio.

same same, really really

No, it depends on what you're going to do with those samples, while it's true that sampling at twice nyquist will give you all the information you need, in some cases you would want to interpolate the values before using them... so you might aswell have sampled at a higher rate in the first place (they're talking about pretty slow signals, where constraints of accurate sample clocks etc aren't an issue.

As I said, it's not the best presented article, it leaves out a lot of information which allows for misconceptions and misunderstandings.

[Jon Hodgson]

Originally Posted by karyoky****

maybe the maxim glossary is more interesting or helpful?

ADC and DAC Glossary - Maxim

It's a useful glossary.