What 2 Channel A/D Is More Transparent Than A Lavry Gold?

[jayfrigo]

Originally Posted by jamsmith

Uh, actually once you are in the digital domain, I don't have any problems there. (I have been a software engineer for 12 years.) But the alegory here doesn't work. Four individual samples are not "4". They individual 4 samples and not added. The math works that way counting the number of bits in each sample where 1+1+1+1 = 0100. And while digital audio is not a "graphics file" it can be represented graphically. When you see a waveform in your editor, you are seen a graphic representation of a sequence of samples. In a sense, a graphic file is not to different as a graphic file is collection of bytes or words that describe individual pixels much as samples describe points on a wave.

As to the sine components and Nyquist, that is what I always knew to be true. It just appeared from the way I read things here (sometime you guys can be brief to the extreme) that something else was believed. I will not elaborate on that, I will accept it as my misinterpretation. I just wanted to clarify it for myself.

Addition like that indeed has nothing to do with actual digital audio, but it's a simple metaphor to illustrate a point. If you really want to run the real math and work the proof for yourself, follow the link Dave Collins gave you. The original Shannon proof and more recent and more complete math is in there IIRC.

And while you can represent the audio graphically, it's really not the same thing. You are not simply identifying points on a wave where more points equals better precision (within the bandwidth of course, it's always about the bandwidth). You are just storing enough information to know how to reconstruct the wave. It's more like vector than bitmap. I'm reluctant to even say that, however, because graphics comparisons are a bad idea for audio, and it's really not vector. Again, just an illustration, the same way points on the image of a wave are an illustration and not completely telling of how it works.

[masteringhouse]

This may help:

Bores Signal Processing - Introduction to DSP - basics: antialiasing

[LNerell]

Originally Posted by jamsmith

Looking at this diagram:

This diagram shows me one simple waveform and two complex waveforms. It seems to me that you are attempting treat all three as the same, a simple waveform without taking into account the other two are made up of multiply waveforms. I suggest along with Bob Katz's book you also pick up a Physics of Sound book as well. Sorry, not trying to be insulting just helpful.

[jamsmith]

Originally Posted by LNerell

This diagram shows me one simple waveform and two complex waveforms. It seems to me that you are attempting treat all three as the same, a simple waveform without taking into account the other two are made up of multiply waveforms. I suggest along with Bob Katz's book you also pick up a Physics of Sound book as well. Sorry, not trying to be insulting just helpful.

Eh, no insult taken. I thought perhaps was being conveyed here (and perhaps is - next quote) is that currently used sampling techniques can recreate complex waves by means other than discrete samples.

Originally Posted by jayfrigo

And while you can represent the audio graphically, it's really not the same thing. You are not simply identifying points on a wave where more points equals better precision (within the bandwidth of course, it's always about the bandwidth). You are just storing enough information to know how to reconstruct the wave. It's more like vector than bitmap. I'm reluctant to even say that, however, because graphics comparisons are a bad idea for audio, and it's really not vector.

Now we are back to square one. Years back when I was first studying digital to analog conversions, what was being done was every nth of a second the instantaneous voltage was measured based on an absolote voltage level and assigned a value. So, for example, in a 16-bit/44.1 converter, 5 volts would be divide in to 65536 discrete steps. For simplicity, assume these steps are linear, a 2.5 volt measurement was stored as 32768. Or 0x8000. So 44100 time per second, a measurement it made and the value stored in a series of 16 bit words. These word where then converted back to an analog signal reversing the process giving you a set of staircased voltage the were filtered to recreated the sinusiodal nature. (Funny that you mention vector graphics because at the time I was learning about vector graphics and wondered why something like that wasn't being used to for recording)

Now here is the disconnect. Are you say that that style of sampling is not what is being used? That rather a set of snapshot of the waveform, we are now doing a mathematical representation of the wave?

Originally Posted by jayfrigo

Again, just an illustration, the same way points on the image of a wave are an illustration and not completely telling of how it works.

What better represents a waveform than an oscilliscope? Sure its a composite for a number of smaller waveforms. But when when we record sound using analog equipment, we are only recording those sums and difference to tape or vinyl. Obviously you can't see it on tape, but if you magnify vinyl, you see the same thing as you do on the scope. It the whole story. Except on vinyl it truly is continuous. An oscope is actually more analogous to a digital display. The line you see on display is created by bombarding the screen with an electron beam - individual electrons that combine to form what your eye percieves as a solid line. Now I doubt you couldn't magnify the oscope screen to view the individual electrons (yes, physics people - it photons created by the bombardment you actually see), you can take a waveform on the computer that looks "exactly" like a snapshot of the oscope screen a blow it up until you see the indivual sample points. Thanks for the book reference. I know a lot about sound already. Real sound in a room is very complex. But once the sound is converted via a transducer, it become no more than varying voltage and it is no more complex than that until another transducer converts it back to sound (or you try to electronically manipulate it - and even that is governed by simple equations for analog equipment).

[Dan Lavry]

Originally Posted by jamsmith

Ok say you are sampling at 44.1 at 20K, this is barely more than 2 samples per period. You would not be able to tell if the sample was sine wave or a triangle or a sawtooth.

First, already mentioned, if 20KHz is your limit, then a square or a sine wave would sound the same shape. But I would agree that the amplitude of the fundumental (square vs sine or tringle) would be different, thus that answer is not complete. So here is what I say:

You are making the same mistake that many others do, and I would suggest that you do read Nyquist works. And for people that do not like math, I did place a document on my web, a 32 page pdf with a lot of plots, and no math. The name of the paper is Sampling Theory. It will clear many misconceptions.

But I will now address your misconception about 2 points in a cycle. True, I could not take 2 sample points and tell you what the signal is. But the fact is that we are not "looking at 2 points". We are "looking at" very many points! Nyquist does not say that we can recover the signal if we sample at twice the highest harmonic. Nyquist says we need to sample slightly twice as fast as the higher harmonic. If you do so, then the previous cycle will have the samples at slightly different sample values *different time locations with respect to the cycle time position, and therefore also different amplitudes. Now go and visit what happened 2 cycles ago, and again you find different values...

Now take a digital wave sampled at say 20KHz, at a rate of say 44.1KHz. Indeed, it does not resemble a 20KHz sine wave. It looks like a real mess, relative to a sine wave, and visually it is difficult to see what is going on. But now put an analog filter capable of removing the high frequency above Nyquist, and what is left is a clean 20KHz sine wave. The filter in fact gets to "connect the dots" into the appropriate shape. It DOES take into account the many previous sample values of the previous cycles. True, the more recent cycles account for more then what happened say 1000 sample ago, however the accumulated impact of all the previous samples is an integral part of generating the proper wave.

The next "stumbling block" in understanding the issue is the fact that it is difficult to make a great filter that will pass 20KHz and block everything above 22KHz (Nyquist). That was solved long ago with the concept of over sampling, which enables the data to be up sampled at much higher rates. Up samping does not add content to the audio. It "fills in" a lot of the "in between dots", thus it does some of the task in the digital domain. A 20KHz wave up sampled from say 44.1KHz to say 1MHz (and most up samplers these days go further) does look a lot more like a 20KHz sine wave. That digital up sampling process does work and the reason again is the fact that earlier samples do contribute to the signal reconstruction, because the values of the samples in previous cycles are different then the 2 points in the last cycle you are looking at.

In fact, that whole business is much more complex. To start with, music is NOT made of equal and repetitive sine waves. I am talking about that concept at as low a level as I can, trying to make some sense for a wide range of readers here. The "natural" follow-up question could be how many past sample does one need to properly reconstruct music, and believe me, the answer will take an EE a year of high level of technical background.

For starters, I will present you with a question:

Given a tone with 1 second duration, say beginning with an "A 440Hz", with a linear pitch bend ending with "A 880Hz", what is the pitch of such tone? It is not 440Hz, it is not 880Hz, nor is it 660Hz (the average). Nor does the ear hear a cycle at a time. Once you answer this well, I will be glad to continue the discussion beyond what is so common in audio, which is an inappropriate view of musical signals. Fourier series says that all PERIODIC waves can be looked at as a sum of simple sine and cosine waves that are harmonically related (by simple integer ratio). Such a tool is great at simplifying life. It enables simple calculations of capacitor and inductor impedance, it makes life very easy, while yielding tons of wrong notions and misconceptions. This level of understanding is somewhat useful, and what I would expect my techs to be familiar with it. The fact is that we do have tools that take us beyond the limitations of periodic waveforms, but those tools require some serious math background. The recently popular phrase "an inconvenient truth" comes to mind. The ear people should use their ears, the engineers should be technical. I happened to be a musician with technical background so I get to see how much is being missed by both sides.

Regards
Dan Lavry
Lavry Engineering, Inc

[jamsmith]

Originally Posted by Dan Lavry

For starters, I will present you with a question:

Given a tone with 1 second duration, say beginning with an "A 440Hz", with a linear pitch bend ending with "A 880Hz", what is the pitch of such tone? It is not 440Hz, it is not 880Hz, nor is it 660Hz (the average).

It not a "fair" question. You gave frequencies and asked for the pitch. They are not the same. Pitch is the perception of a note and it not always the same as the frequency indicates it would. Assuming you mean the resulting tone, ideally the answer would be A5. In the real world, we identify it as slightly flat because you have only had the 880Hz for a single cycle and no human ear could detect that. Further, perception of pitch can also be a factor of amplitude that you have not given here. And if we disregard the musical terms and apply scientific terms, the result would not be expressed in Hertz but in mels. In this case: ~858 mels.

Originally Posted by Dan Lavry

the answer will take an EE a year of high level of technical background.

You are greatly respected here and I mean no disrespect, but when I decided I wanted to be a software engineer with no degree, no training and no more than a book on BASIC and C primer on my shelf, I heard a lot of that. It was less than a half a year later that I was employed fixing bad programming done by Ph.Ds. You may have noticed I put forth a lot of "notions". I find in the end, I get a lot more useful information that way. The ancients often taught via debate. I can post a question on a forum and get natch or a handful of half-assed answers. But post a "misconception?, and you can end up with 4 or 5 pages of responsed! I swear the more ludicrous the supposition, the greater the response!

And not only do I mean no disrespect, I am grateful someone of your caliber will suffer fools long enough to dispense some wisdom. The level of participation by top professionals is the reason I am here.

[jayfrigo]

Originally Posted by jamsmith

Now here is the disconnect. Are you say that that style of sampling is not what is being used? That rather a set of snapshot of the waveform, we are now doing a mathematical representation of the wave?

Couple responses of a limited nature, and then we've probably taken this as far as it can go. We are definitely using discrete samples, but the math component can't be ignored. It's not simply a graphics file of the wave we are making. Check out that Dave Collins link several posts above. Remember calculus class? How do you find the slope at an instantaneous point for a curving line? You take two points and cut it up into smaller and smaller distances etc. And how do you calculate where it is going?

The math informs the bit of the puzzle I think you are questioning. I can't go too far into it because frankly I'm not qualified (but Dan is!). I teach bachelor level applied physics of a limited nature, not doctoral math, so I only know enough heavy math to be dangerous. But I have tinkered with the original math and it's pretty cool. It's right in that link. Have a go at it. It's actually kind of fun. OK, I guess I'm just a geek...

Yes, there are points, quantization, filters, smoothing, all the stuff you get already. Yes, the square doesn't look very square at >20kHz because everything above Nyquist is missing. That doesn't mean the <Nyquist stuff is any less there. The point this began with was that the in-band stuff is not improved with the higher sample rate. The "square edges" above Nyquist are not in-band.

Also, to be precise, analog tape and vinyl records do have limitations physically. There is no such thing as infinite resolution, and tape has domains that are arranged by polarity, and there are a finite number of them. You could probably figure out what the digital equivalent resolution was if you had the time and the inclination.

Aso, I think Dan was referring not to a particular pitch, frequency, or value in mels, but the point was rather that a bending note has no steady pitch. The ever changing nature of the sound was his point if I understood him correctly, and that's where the math helps you to appreciate what's going on, even if you only dig part way into it.

OK, I think that about does it, at least for me. I think everybody is close enough to the same page now, and running over the same ground in different ways is probably just going to muddy things up rather than making them clearer at this point.

[Sui_City]

Jay and all others,

I found this a few minutes ago and thought that it would just show that the high minds of audio are not the only ones realising that higher rates are not better.

Here is a link to an article by a camera testing lab stating that 6 megapixels is the optimal resolution for small cameras.

» Best picture quality with 6 megapixels!

Sorry for the tangent, but given that the "higher resolution image" analogy is brought up so often, i thought this would be an appropriate addition to the discussion.

[Peter Oxford]

What this article claims is true, but only holds true for compact cameras utilising a small sensor.

In other words, yes this is right, but there may be other factors you need to take into account in consideration of image capture, that mean these facts are not faithfully extrapolatable as an absolute truth throughout the entirety of the photographic universe.


Photography is really all about understanding, appreciating, capturing and manipulating the varying quality of light. Knowing how to use it creatively, and faithfully reproduce that image within the absolute confines and severe limitations of a heavily constraining medium.


Like sound.

It’s at the extremes of the range that things become difficult to get right really.

Our eyes once properly accustomed to the available light are capable of seeing detail in darkness that a Camera will find impossible to replicate.

At the bright end of the scale, replicating a stunningly bright light without whiting out in a manner that betrays and beggar’s all notion of quality, creates equal difficulties.

Our eyes like our ears, can tell whether something is being reproduced correctly or incorrectly, most especially, at the extreme’s where authentic reproduction is extremely difficult and the limitations of the bandwidth of the carrying medium come heavily into play.

Provided we have the intimate experience of knowing what ‘correct’ is really and truly like.

It’s what photographers refer to as ‘Dynamic Range’ and I suppose it’s a bit like ‘frequency response’ in audio.


I believe Fuji were the first manufacturer of Compacts to level off at 6 MP.

That doesn’t mean their cameras have leveled off there, as the pixel count on modern cameras has continued to escalate, however, the thing is, they are using the extra pixels to deal specifically with the ‘difficult to reproduce faithfully’ areas of the image.

Here’s some support to your hypothesis from Fuji, which also explains what I mean, hopefully with greater clarity, in regard to their better quality cameras.


“Sensor

The ultimate goal in development of Fujifilm's D-SLR cameras has been to maximise image quality through constant R & D into the Company's proprietary sensors.

With the introduction six years ago of the acclaimed Super CCD, Fujifilm has moved away from increasing the pixel density in favour of image quality developments that increase the overall performance of the sensor.

The FinePix S5 Pro will see the refinement of a winning formula - Fujifilm's Super CCD SR II will be updated to the new Super CCD SR Pro.

Using a unique layout of twelve million paired photodiodes (6.17 million larger 'S' photodiodes for main image information, combined with 6.17 million smaller 'R' photodiodes for bright area information), the S5 Pro will deliver improvements in noise, dynamic range, colour and tonality.

Further improving the capability of the sensor, a new, improved low-pass filter will ensure that moiré and noise are kept to an absolute minimum.

Fuji technicians believe improvements in these key areas will be of more true value to professional photographers - the challenge is quality of information, not quantity of information.


Dynamic range

The FinePix S5 Pro will also afford the user much greater control over the unique image parameters that its technology provides. Six stepped dynamic range settings between 100% (normal dynamic range) and 400% (dynamic range expanded by two stops), will enable the photographer to achieve exactly the tonal feel to their images that they want.

In addition, three distinct settings for the 'negative film' preset option will further enhance the customisable nature of the camera, giving portrait photographers precisely the 'look' that they wish to create.”



To me, this is an interesting solution, because they haven’t denied the problems that exist, but taken them on board, developing and utilising copious amounts of resources to provide specific and convincing solutions to the areas of difficulty.

Perhaps we could liken this in audio to up-sampling where it is is rife to solve a great many issues that might otherwise be a cause for concern.

Therefore I believe their assertion “the challenge is quality of information, not quantity of information”

To be somewhat misleading, and entirely challengeable in itself.


Clearly, we need both quality in the right amounts and quantity in the right amounts for the application in progress.

That 6M Pixel picture, enlarged for use as an advertising poster or billboard, would be utterly inferior to anything we would be used to seeing, and simply no where good enough to use.



So we have professional solutions for professional applications.

35mm type and Medium Format Digital Cameras exist at a professional level which are of extremely high quality and capable of pixels counts well in excess of twenty millions.



Everything is relative...

As the article we were directed to itself postulate’s.


“Digital SLRs with many pixels are okay.

Digital single lens reflection cameras (SLRs) basically show the same behaviour but the sensor of those cameras and the single pixel is much bigger.

Therefore, the cameras have higher sensitivities and show less noise. The high quality lenses provide the necessary resolution and the cameras are designed for high pixel counts.”


Camera Lenses are available for certain makes of Digital Camera which are designed for an optical resolution capable of delivering to the camera back’s sensor 30 + MP.

These cameras are capable of seeing deeper into the shade and varying degrees of light whilst retaining more depth of colour and a higher image contrast ratio. Those characteristics are an essential part of any high calibre reproduction.


As I said....

“We need both quality and quantity in the right amounts for the application in progress.”

The thing is certain people will be far more demanding of higher quality than others.



I believe we were discussing’ high end’ sound reproduction......

As a footnote, I don’t have any problem at all with Mr. Lavry’s white paper, or his assertion that 60 kHz is an optimal sample rate.

I never have done.

The issue for me is do we assume everything else is equal?

If it is not, is there a qualitative improvement available somewhere to be had because of this?

In other words, do we assume there is a level playing field throughout a design, and the implementation of every other part, as faithfully detailed in execution as Mr. Lavry’s white paper?

The fact is, in manufacturing audio designs, shortcuts and trade off’s often occur that compromise and violate the strict principles loftily espoused in such white papers. Most particularly, the quality of implementation at the analogue stage and the filtering above 20 kHz are of particular concern.

In other words, my view is that it is not simply a matter of getting a particular factor correct like sampling rates for instance, but a great many factors, that are all closely interrelated, and unsurprisingly, can all directly affect each other. Every aspect of the entire design needs to be given the consideration that their importance requires.

If everything else is not equal, (or theroetically ideal) and shortcuts elsewhere in the design have been taken, then we are in a game of trade off’s and compromise for practical considerations which include cost as a vitally important, determinative factor.

Here is where the difference between ‘theoretically ideal’ postulation, and cost effective ‘practical implementation’ give birth to a divergent experience for users, resulting in corresponding argument.

And that, as far as I am concerned, is really why the issue of sample rates becomes a ‘live point’ in discussion.

And why, irregardless, I would always use the available solution, that sounds best to my ears.

Irrespective of sampling rate or any such detail of specification.

Quite regardless.






P

[Sui_City]

Peter,

The reason for my posting a link to that article was purely to illustrate to some that the most simplistic argument, (more automatically equals better), is not always the case.

And as you rightly point out, the interactivity of the many different "factors" within a design plays a huge part.

I just wanted to add it as another example of the counter-intuitive.

[Dan Lavry]

Originally Posted by jamsmith

It not a "fair" question. You gave frequencies and asked for the pitch. They are not the same. Pitch is the perception of a note and it not always the same as the frequency indicates it would....

You are greatly respected here and I mean no disrespect, but when I decided I wanted to be a software engineer with no degree...

And not only do I mean no disrespect, I am grateful someone of your caliber will suffer fools long enough to dispense some wisdom. The level of participation by top professionals is the reason I am here.

Hi,

First, I too mean no disrespect. I never said that one must have a Phd or any degree to acquire knowledge. I said that the math required to deal with the issue is pretty high level stuff.

Now back to the point. I though I did open the door to provide some understanding why the view of "2 point in one cycle" yields to a confusion. I was tempted to put a plot of 10 cycles of 20KHz at 44.1KHz sampling, to show how the points keep "marching" on the wave, in such a way as to "define" different values on each previous cycle. But I figured you can do it yourself.

You call my question "unfair"? The reason I asked it is to stimulate thought about the subject. If you can answer the question I asked, you will have a better grasp.
My question about continuously variable frequency is a very good learning tool. Clearly the sound of a smooth pitch bend is a smooth pitch bend, and the recording process works, though it does not rely on the rather simplistic case of forever repeating sine waves (periodic signals). In fact, in my question, each "new cycle time" (time between zero crossings) of the sine wave is shorter, so it is easier to see that you are defining different sample points, you do not even need to plot it...

Your answer is far from what I expected so far. You say I gave frequencies and asked for pitch.
First, when we talk about periodic waves, and in my question I was using sine waves. The pitch is the "frerquency of the lowest harmonic" and with a sine wave all we have is the lowest harmonic. But in my question, I was asking what is the frequency (or the pitch) of a 1 second duration tone with a pitch bend, where the first cycle is long duration (corresponding to A440), and the last cycle is short duration.

You said that pitch is a perception of a note, and is not always the frequency. Maybe so, maybe not but that is not at all the point. The wave I described is not limited to observation and perception of the ear. We can plot it on a page, we can see it on a scope, we can generate air pressure variations (sound) that starts with slower vibrations and speeds up over 1 second.

There is a lot that can be learned from steady state analysis of periodic waves, and that technology comes in handy for all sorts of testing and debugging. But understanding of Nyquist and sampling requires a whole different view. The question of "tracking" a signal with "relatively few points" requires deeper understanding. The first step, in my view, is to let go of some of those fundamentals that do not apply and only "drug you down". A 20KHz constant signal is great for checking the speed of some circuit (slew rate), but not for explaining how sampling works. For that, we need to be thinking about ANY signal.

And of course, there is no way we can sample and reconstruct ANY signal, however, we can have perfect results if we can put some RESTRICTIONS on the sample signal and similar restrictions on the reconstructed signal. The fact that we pre filter a signal before sampling, changes the signal from "anything can happen" to "there are limitations as to what can happen". A filtered signal can not change faster then what the pre filter allows. No more "real fast" wiggles in the wave form... Of course when you sample, you introduce a lot of "sudden jumps" but all that happens at frequencies higher then the filtered signal. The post filtering (after the DA) removes the high frequencies, leaving the original wave-shape of the pre filtered signal.

It is not easy or intuitive to see how the post filter has one and only one way to "connect the dots", and that way yields the original shape. But as I pointed out, the previous sample values play a role in "steering" the reconstruction filter. The more recent samples play a bigger role then say 100 samples ago (which is about 2.2 msec into the past). Clearly, to have infinite accuracy, we would need to look way back into the past, to a time way before Moses crossed the sea... But say we want to have a part per million accuracy, then a few msec may be enough.

Again, the easiest way I know to prove that Nyquist works is by use of higher level math. Intuition and street smarts does not always work, and can yield wrong results. It could be helpful for people to realize that digital audio is only a "thin slice" in the subject of conversion, which covers a wide range of medical applications, instrumentation, industrial control, video, telecommunication.... If Nyquist were wrong, we would not have much of the technology we are so accustomed to. Conversion is not just for the ear. True, conversion for audio is different then for say video, but here we are talking about basic principles that cover all of conversion.

Again, it may be helpful to read "Sampling Theory" on my web.

It would be good to get some feedback: Are my comments helpfull?

Regards
Dan Lavry

[jayfrigo]

Originally Posted by Dan Lavry

It would be good to get some feedback: Are my comments helpfull?

Regards
Dan Lavry

I can't speak for others, but I think your contributions are very helpful to the thread. Thank you for taking the time.

[jamsmith]

Originally Posted by Dan Lavry

Hi,
It would be good to get some feedback: Are my comments helpfull?

Abolutely. Not so much directly as it is obvious that it far to complex a subject to cover in a post. But I have a flavor for where I need to expand my understanding. And during the interim of the discussion I have located and read a lot more on the subject. And I will be reading your article after this post. However, while the explanations and the math for the theory is ubiquitous in literature, a few reference for the application of the theory would be nice. Schematics (exempler, rather than actual) of these components used in building these devices go as long a way to understanding as do formulas.

Originally Posted by Dan Lavry

You said that pitch is a perception of a note, and is not always the frequency. Maybe so, maybe not...

Well, if you go by standard dictionary definitions, pitch and frequency are identical. However, in all my references on acoustics, music theory, the physics of musical instruments, etc, all define pitch specifically as "the perception of music. Based on the work of Stevens and Volkman.

But in all fairness to you, the discussion was in the realm of electronics where pitch and frequency are interchangeable as in "pitch-to-voltage converter". As I mentioned, I am just getting back into electronics after a decade.


Maybe one last question if I may. The conclusion from the experts' posting is that (allowing for dynamic range) a 16/bit 44.1kHz converter should be all we need. So where do the 24bit/96kHz converters fit into all this? It is just that these units are built with better filters, etc?

[DSD_Mastering]

Originally Posted by jayfrigo

I can't speak for others, but I think your contributions are very helpful to the thread. Thank you for taking the time.

I second this!thumbsup
I sat with Dan at the AES convention and probably took up about an hour of his time just talking about being musicians and why he won't let someone record his piano playing.
Come on over to my studio any time Dan!

Regards,
Bruce

[jayfrigo]

Originally Posted by jamsmith

Maybe one last question if I may. The conclusion from the experts' posting is that (allowing for dynamic range) a 16/bit 44.1kHz converter should be all we need. So where do the 24bit/96kHz converters fit into all this? It is just that these units are built with better filters, etc?

Well, I don't think that's quite what is being said. 64 kHz / 20 bit I think would probably be enough for a delivery format. Processing, especially non-linear, has reason to use even more. And oversampling for conversion still is useful. Filter issues are indeed a big deal when it comes to 44.1. There are plenty of reasons to use more than 44.1/16, but more detailed representation of stuff well below Nyquist is not one of them, nor do I believe that the ability to record low-level ultrasonic harmonics is one either, nor is the sky the limit as to how much more resolution is helpful before the tradeoffs become too great.

[bob katz]

Originally Posted by jayfrigo

Well, I don't think that's quite what is being said. 64 kHz / 20 bit I think would probably be enough for a delivery format. Processing, especially non-linear, has reason to use even more. And oversampling for conversion still is useful. Filter issues are indeed a big deal when it comes to 44.1. There are plenty of reasons to use more than 44.1/16, but more detailed representation of stuff well below Nyquist is not one of them, nor do I believe that the ability to record low-level ultrasonic harmonics is one either, nor is the sky the limit as to how much more resolution is helpful before the tradeoffs become too great.

What I like about you Jay, is you talk the talk, accurately and succinctly!

BK

[Lerxst]

Originally Posted by jayfrigo

I can't speak for others, but I think your contributions are very helpful to the thread. Thank you for taking the time.

I will post just to say that I agree that all of this discussion is very much appreciated. I am just reading and learning, and Dan, Bob, Mike, Jay etc... Please continiue...

People are watching.

~Lerxst

[silverking]

Dan, you are one of about 5 people on the internet who's posts I take as de facto information on the topic under discussion.

So, that's a big "YES"...... most helpful.

I find it amazing that folks at the top of their game even bother to post. The signal to noise goes up substantially when a big name shares their ideas.

Personally, I thank you for your time and thoughts.

[masteringhouse]

Absolutely helpful Dan!

There is one question I still have though as far as encoding a signal. It appears from your discussion and this link:

Bores Signal Processing - Introduction to DSP - basics: antialiasing

that there can be frequency-dependent amplitude modulation superimposed on the original frequency(ies) being encoded using Nyquist theory alone. Is this usually corrected in the converter before encoding? If not, how do outboard digital processors and plug-ins deal with this since it seems that this can be problematic for functions like limiting? Or is this another reason for upsampling before limiting via digital only?

[Nordenstam]

Originally Posted by jamsmith

However, while the explanations and the math for the theory is ubiquitous in literature, a few reference for the application of the theory would be nice. Schematics (exempler, rather than actual) of these components used in building these devices go as long a way to understanding as do formulas.

This application note is from 1980 and shows block diagram of hardware and an outline of the big ideas: http://www.national.com/an/AN/AN-236.pdf

There's also a lot of (practical) application notes to be found on the websites of just about every semiconductor manufacturers with such devices.


Andreas N

[Dan Lavry]

Originally Posted by masteringhouse

Absolutely helpful Dan!

There is one question I still have though as far as encoding a signal. It appears from your discussion and this link:

Bores Signal Processing - Introduction to DSP - basics: antialiasing

that there can be frequency-dependent amplitude modulation superimposed on the original frequency(ies) being encoded using Nyquist theory alone. Is this usually corrected in the converter before encoding? If not, how do outboard digital processors and plug-ins deal with this since it seems that this can be problematic for functions like limiting? Or is this another reason for upsampling before limiting via digital only?

Well, I would expect a practical converter operating at say 44.1KHz to yield a clean response to a little over 20Khz. But Nyquist for 44.1KHz is at 22050Hz, and the performance at say 22049Hz will indeed be terrible. This does not mean that Nyquist was wrong. So where is the discrepancy?

If you push the input signal to say 1Hz below Nyquist, there is a "beat" of 1 second. In other words, the input signal and Nyquist get out of phase and back into phase once per sample. The waveform looks terrible and it is terrible. So what kind of a filter would I need to "clean it up"? Clearly I can not do much with say 1000 tap FIR filter, because a 1000 tap filter only look at the last 0.024 seconds. In fact, to look at one second of data I would need 44100 taps, a huge filter, and even that is not long enough to offer enough correction. So pretty soon we end up with samples going back to the time when Moses crossed the sea... I was using taps and FIR. If you want to talk about analog filters, you would need thousands of opamps and many thousands of parts with precision far beond praticality. I other words, we are pushing the practical requirements too hard towards the theoretical limits. That in turn, calls for near theoretical filters.

But instead of trying to sample and reconstruct to within a few cycles of Nyquist, we used to filter the energy as much as we can with enough margin away from Nyquist, to enable the use of practical filtering. And these days (the last 15-20 years), with over sampling AD front ends and up sampling DA at the back end, the signal frequncies are so far from the LOCALIZED rate (Localized Nyquist, which is not the data sample rate, big distinction worth keeping in mind), thus the issue is completely gone for all practical purposes.

The theory is correct. The "demonstration" of an unfiltered signal is misleading. In fact, it takes a lot of computer resources to simulate a near brick wall filter at say 22.000Hz (50Hz away from Nyquist) and run say 10 seconds of signal through it. But as you improve the simulation, the "beat effect" starts to disappear.

Regards
Dan Lavry

[masteringhouse]

Thanks Dan!

This cleans up the loose ends (no pun intended) for me.

[matyas]

These last few posts really clarified a lot of things for me. Thanks, everyone. I'm still trying to wrap my head around a few things, so bear with me a moment.
It really helped to think in terms of the "squaring" of waveform which occurs during the A-D process as adding high-frequency information which will be removed by filtering. (At least, that's how I was understanding some of this stuff.) I had known for years that A-D converters need a steep analog filter a little below the Nyquist frequency to prevent aliasing, but considering the distortion of the analog waveform which occurs during A-D as an addition of high-frequency artifacts reinforces in my mind the crucial role of the analog filter in a converter's sound. It sounds as if the quality of filtering would be one of the most important components when people talk about a converter's "analog section". (I do realize that there are other factors involved in defining a converter's sound, of course.)
Also, this makes me wonder if this is a contributing factor to why so many people like to use tubes, or transformers, or even analog tape before converting to digital - in other words, to smoothly roll off the high end prior to conversion. It also makes me wonder if the current trend back towards mixing analog, or at least using highly colored analog gear during mixing, has something to do with "smoothing out" high-frequency components added during A-D and then not filtered out during D-A?