Struggling to understand digital audio.
Clueless
As per subject, I'm struggling with several aspects of digital audio theory, and I think you will able to illuminate me. Please jump in and correct me at any time, because I must be going wrong somewhere, and I want to know.
I think of part of Nyquist as saying "You must have *more* than two samples per cycle of the highest frequency you want to record". (This is easier for my brain, but will thinking of it like this trap me later on?)
So ideally, there would not be any information at the Nyquist frequency or above in the input. OK fine, no problem understanding that.
Now imagine that the input filter had intentionally been left out, and a signal of exactly the Nyquist frequency was deliberately injected. This signal would not be recorded properly. It is possible that the sampling window would coincide with the zero crossing points of the signal, or any other point on the signal, extremely unlikely to be the -ve & +ve peaks, right?
Then remove the first signal, and inject a signal of a frequency very close to, but just slightly less than the Nyquist frequency. This signal would then be being sampled at a rate slightly higher than 2 samples/cycle.
Here's my problem - I think that the recording of this signal would beat. Sometimes the sampling window would coincide with the peaks of the signal, and sometimes it would coincide with the zero crossing points. How is this corrected later?
So when it is said that the input must be bandwidth limited to frequencies *less* than the Nyquist frequency, how much less is *less*? To me, it seems that the nearer an input frequency approaches the Nyquist frequency, the more error is possible. But this goes against many assertions I have read about sampling, so what fundamental am I missing?
Alright then, dither. I find this extremely counter-intuitive, but I've given my intuition a smack on the head, and let's assume I've seen Ken Pohlman's explanation and accepted it.
Given that a properly dithered signal completely removes all quantisation distortion, the advantage of longer word converters is then potentially lower noise floor, greater dynamic range, and *not* greater resolution? (Plenty of potential for error in that sentence!)
I've read that full scale on a 16 bit converter is conventionally 2V. Is that right? What is full scale on a 20 bit or 24 bit converter? Have we divided the scale more finely, or enlarged the scale, or both?
Can you recommend a properly dithered CD that fades away to silence without getting grainy for me to listen to?
CD distortion at low levels: What's the story? IIRC Noel Keywood(?) tests equipment for one of the UK audiophool magazines. He routinely reports that CD players have 30% THD(?) at -60dB or similar (I may well have remembered that incorrectly). How's he making this measurement? What does it mean?
I can't escape the feeling I'm overlooking things very basic & embarrassing. Anyway, thank you for your help,
Yours Cluelessly.
_______________________________________________
Submitted via WebNewsReader of http://www.interbulletin.com
Michael R. Kesti
>As per subject, I'm struggling with several aspects of digital audio theory, and I think you will able to illuminate me. Please jump in and correct me at any time, because I must be going wrong somewhere, and I want to know.
>
>I think of part of Nyquist as saying "You must have *more* than two samples per cycle of the highest frequency you want to record". (This is easier for my brain, but will thinking of it like this trap me later on?)
If that's easier for you, then that's just fine, and there is no trap.
Frequency and period are inverses, and you have found another
consequence of that.
>So ideally, there would not be any information at the Nyquist frequency or above in the input. OK fine, no problem understanding that.
>
>Now imagine that the input filter had intentionally been left out, and a signal of exactly the Nyquist frequency was deliberately injected. This signal would not be recorded properly. It is possible that the sampling window would coincide with the zero crossing points of the signal, or any other point on the signal, extremely unlikely to be the -ve & +ve peaks, right?
Unless the input signal is synchronized to the sample rate, the
probabilities of sampling at any specific phase are equal.
>Then remove the first signal, and inject a signal of a frequency very close to, but just slightly less than the Nyquist frequency. This signal would then be being sampled at a rate slightly higher than 2 samples/cycle.
>Here's my problem - I think that the recording of this signal would beat. Sometimes the sampling window would coincide with the peaks of the signal, and sometimes it would coincide with the zero crossing points. How is this corrected later?
There is no beat at this or any other frequency, as no non-linear
(nor linear!) mixing of the frequencies occurs. The "correction"
is accomplished by the reconstruction filter, which does the same
job at this frequency as it does at any other.
>So when it is said that the input must be bandwidth limited to frequencies *less* than the Nyquist frequency, how much less is *less*? To me, it seems that the nearer an input frequency approaches the Nyquist frequency, the more error is possible. But this goes against many assertions I have read about sampling, so what fundamental am I missing?
The limit is a function of the input filter. If you could build a
filter that is flat to 22.04 KHz and down 80 dB at 22.05 KHz, you could
go out to 22.04 KHz in a 44.1 KHz sampling system with acceptible
performance. But you can't have that filter in the real world, so you
wind up with a bit less bandwidth with realizable filters.
>Alright then, dither. I find this extremely counter-intuitive, but I've given my intuition a smack on the head, and let's assume I've seen Ken Pohlman's explanation and accepted it.
>Given that a properly dithered signal completely removes all quantisation distortion, the advantage of longer word converters is then potentially lower noise floor, greater dynamic range, and *not* greater resolution? (Plenty of potential for error in that sentence!)
Resolution, noise floor, and dynamic range are all the same thing in the
digital domain.
>I've read that full scale on a 16 bit converter is conventionally 2V. Is that right? What is full scale on a 20 bit or 24 bit converter? Have we divided the scale more finely, or enlarged the scale, or both?
I'm unfamiliar with any such conventions.
>Can you recommend a properly dithered CD that fades away to silence without getting grainy for me to listen to?
>
>CD distortion at low levels: What's the story? IIRC Noel Keywood(?) tests equipment for one of the UK audiophool magazines. He routinely reports that CD players have 30% THD(?) at -60dB or similar (I may well have remembered that incorrectly). How's he making this measurement? What does it mean?
Sorry, I have no answers for these.
>I can't escape the feeling I'm overlooking things very basic & embarrassing. Anyway, thank you for your help,
--
========================================================================
Michael Kesti | "And like, one and one don't make
| two, one and one make one."
mke...@gv.net | - The Who, Bargain
Stewart Pinkerton
>Clueless wrote:
Most of the questions were well answered by Mr Kesti.
>>Can you recommend a properly dithered CD that fades away to silence without getting grainy for me to listen to?
Try Holsts's Planets, the Karajan or Previn versions. The fade down of
the choir at the end of Neptune should descend smoothly into the noise
floor.
>>CD distortion at low levels: What's the story? IIRC Noel Keywood(?) tests equipment for one of the UK audiophool magazines. He routinely reports that CD players have 30% THD(?) at -60dB or similar (I may well have remembered that incorrectly). How's he making this measurement? What does it mean?
This would be more common for -90dB (it's typically 2% or so at
-60dB), but there is a *very* important point to be made about such
measurements. They are not just THD measurements, they are THD+N
measurements, including the noise floor. Given that the noise floor of
a properly dithered CD is 93dB below peak level, it should be obvious
that a *perfect* sine wave at -90dB will show a THD+N measurement of
about 30%! With the best DACs such as the dCS 'RingDAC' in the Arcam
CD-23 and CD-92, actual distortion products measured with a spectrum
analyser show that the real linearity errors are more like 0.1% at
-60dB and about 10% at -90dB. Note that signals as low as -100dB can
be replayed with less than 0.2dB error, indicating that the basic
converter linearity is even better than these measurements indicate.
While we can agonise over these ultra-low-level effects endlessly,
it's worth remembering that the *noise floor* of vinyl is typically
-60dB or so referenced to peak cutting level in the midband, and a lot
worse at the frequency extremes, so all things are relative!
--
Stewart Pinkerton | Music is art, audio is engineering
Arny Krueger
"Clueless" <donot...@interbulletin.bogus> wrote in message
news:3ABD49B1...@interbulletin.com...
> Alright then, dither. I find this extremely counter-intuitive, but
I've given my intuition a smack on the head, and let's assume I've
seen Ken Pohlman's explanation and accepted it.
> Given that a properly dithered signal completely removes all
quantisation distortion, the advantage of longer word converters is
then potentially lower noise floor, greater dynamic range, and *not*
greater resolution? (Plenty of potential for error in that
sentence!)
In virtually all converters, the effective resolution is set by the
noise level. The dynamic range is set by the noise level. The noise
floor is set by the noise level. Therefore none of them is greater or
less than the other.
> I've read that full scale on a 16 bit converter is conventionally
2V. Is that right?
No, in fact some converters have an output that is in the form of a
current, not a voltage. Full scale can be any voltage that someone
set, if they set it as a voltage. Whoever told you this does not
understand how things work.
>What is full scale on a 20 bit or 24 bit converter?
Full scale is whatever someone chose for whatever reason.
> Have we divided the scale more finely, or enlarged the scale, or
both?
The designer of the converter can choose to do one or the other or
both.
> Can you recommend a properly dithered CD that fades away to silence
without getting grainy for me to listen to?
As a rule, CDs don't get grainy when they fade to silence. There
are lots of reasons. If you want to listen to things getting grainy
when they fade to silence please visit
http://www.pcabx.com/technical/bits44/index.htm .
> CD distortion at low levels: What's the story?
There shouldn't be any, and even cheap midfi CD players don't do it.
> IIRC Noel Keywood(?) tests equipment for one of the UK audiophool
magazines. He routinely reports that CD players have 30% THD(?)
at -60dB or similar (I may well have remembered that incorrectly).
He's either using a test setup that is broken or he's testing serious
junk.
> How's he making this measurement?
What ever he does, IME he's doing it wrong or he's found the mother
lode of sub-hifi junk. Cheap $75 rack systems and $50 boom-boxes (as
opposed to the good ones)?
I've tested over 50 converters and AFAIK I've rarely ever seen
anything like that. Maybe a few but we're talking $19.95 PC sound
cards. Furthermore I think that I may be the world's champion tester
of junk converters - testing converters that are of far lower quality
than anybody would use for hifi. PC sound cards are a highly variable
lot - everything from soup to nuts, from some real trash to some
really great stuff.
Just to set the record right, here is a really good PC sound card:
http://www.pcavtech.com/soundcards/CardDDeluxe/index.htm . It
performs better than any CD player can, no matter what the price.
> What does it mean?
He's doing something strange (I hope).
Here's my collection of converter tests: www.pcavtech.com. I measure
every converter at -60 dB.
Here's an example of one of my tests.
http://www.pcavtech.com/play-rec/Sony_D-220/index.htm#DR_DA . If you
look at the graphic, you will see a number for THD. It says 0.28%.
This is a $60 portable CD player, If cheap CD players had lots of
distortion at low outputs, you'd think that a $60 one would have lots
of distortion at low outputs, no? BTW a modern $25 CD player does
not perform that much worse.
How about looking at some REAL garbage, a totally suzzoid PC sound
card (not a good one). Hmm, one of the most suzzoid sound cards I've
ever tested.... How about this one:
http://www.pcavtech.com/soundcards/als200/index.htm#DR_AD . Well how
about that! We've got a winner folks! 25% THD. OK, this is a real POS
and it sounds bad at any listening level. For a real thrill, look at
the frequency response:
http://www.pcavtech.com/soundcards/als200/index.htm#FR_LB . It sounds
something like a telephone or AM radio. I've done some informal tests
on $25 CD players. They are far better.
How about a 3 year old midfi CD player?
http://www.pcavtech.com/play-rec/CDP-XE_500/index.htm is a player I
picked up for $99 something like 3 years ago as "B" ("refurbished")
stock in a Sony factory store.
http://www.pcavtech.com/play-rec/CDP-XE_500/index.htm#DR_DD tells the
story: 0.044% THD at - 60 dB. No visible harmonics of any kind, just
noise.
André Huisman
> I think of part of Nyquist as saying "You must have *more* than two
samples per cycle of the highest frequency you want to record". (This is
easier for my brain, but will thinking of it like this trap me later on?)
> So ideally, there would not be any information at the Nyquist frequency or
above in the input. OK fine, no problem understanding that.
Not idealy, it's a DEMAND!
> Now imagine that the input filter had intentionally been left out, and a
signal of exactly the Nyquist frequency was deliberately injected. This
signal would not be recorded properly. It is possible that the sampling
window would coincide with the zero crossing points of the signal, or any
other point on the signal, extremely unlikely to be the -ve & +ve peaks,
right?
Yep. The signal would mirror onto itself, resulting in a DC level if it's
exactly 1/2 Fs.
> Then remove the first signal, and inject a signal of a frequency very
close to, but just slightly less than the Nyquist frequency. This signal
would then be being sampled at a rate slightly higher than 2 samples/cycle.
Yep.
> Here's my problem - I think that the recording of this signal would beat.
Sometimes the sampling window would coincide with the peaks of the signal,
and sometimes it would coincide with the zero crossing points. How is this
corrected later?
What you are forgetting is that turning ON a signal requires a substantial
amount of bandwidth. Filter a signal close to FS according to the Nyquist
theorem and you won't see a certain On/Off state. It gradually goes there
(and goes back down again).
> So when it is said that the input must be bandwidth limited to frequencies
*less* than the Nyquist frequency, how much less is *less*?
Infinitely small (theta).
> To me, it seems that the nearer an input frequency approaches the Nyquist
frequency, the more error is possible.
Nope, zero error.
> But this goes against many assertions I have read about sampling, so what
fundamental am I missing?
You're missing the fact that if a signal is (say) 1/2Fs -1Hz, it MUST take
some 10 seconds before it reaches correct level after filtering (and prior
to sampling). Output will behave identical and WILL provide a perfect copy
of the input signal (when looked at AFTER the input filter).
> Alright then, dither. I find this extremely counter-intuitive, but I've
given my intuition a smack on the head, and let's assume I've seen Ken
Pohlman's explanation and accepted it.
> Given that a properly dithered signal completely removes all quantisation
distortion, the advantage of longer word converters is then potentially
lower noise floor, greater dynamic range, and *not* greater resolution?
(Plenty of potential for error in that sentence!)
Analogy: Put fingers in front of eyes: You miss out on a LOT of information
in your sight. Now move hand quickly to left and right while keeping it in
the line of sight. You can see everything! Your fingers are still there,
it's just that their action is sort of "scattered".
> I've read that full scale on a 16 bit converter is conventionally 2V. Is
that right?
CD-players generally have a 2Veff (or 2.8V peak_zero) output level at 0dBFS.
The actual converters can have an output level whatever the designer wanted
(quite often a current output instead of a voltage output BTW).
> What is full scale on a 20 bit or 24 bit converter? Have we divided the
scale more finely, or enlarged the scale, or both?
Same output level for a CD-player containing these converters. Converters
themselves, again, can have any kind of output level (voltage or current).
> Can you recommend a properly dithered CD that fades away to silence
without getting grainy for me to listen to?
Sorry, I mostly listen to the song, not the outro ;-)
> CD distortion at low levels: What's the story? IIRC Noel Keywood(?)
tests equipment for one of the UK audiophool magazines. He routinely
reports that CD players have 30% THD(?) at -60dB or similar (I may well have
remembered that incorrectly). How's he making this measurement?
Residual noise PLUS distortion. Quite probably such measurements are done
prior to the output filter (for whatever hidden agenda). If there is a
substantial amount of signals >1/2Fs at the actual output then clearly there
is something wrong with the device in question (quite probably said device
will also have a glass frontplate, wooden panels and maybe even some
tubes/valves sticking out of it).
--
André Huisman
New Line licht & geluid
hui...@new-line.nl
http://www.new-line.nl
--- pardon my French, I'm Dutch ---
Richard D Pierce
Clueless <donot...@interbulletin.bogus> wrote:
>
>I think of part of Nyquist as saying "You must have *more* than
>two samples per cycle of the highest frequency you want to
>record". (This is easier for my brain, but will thinking of it
>like this trap me later on?)
Nope, this is exactly correct.
>Then remove the first signal, and inject a signal of a
>frequency very close to, but just slightly less than the
>Nyquist frequency. This signal would then be being sampled at
>a rate slightly higher than 2 samples/cycle.
>Here's my problem - I think that the recording of this signal
>would beat. Sometimes the sampling window would coincide with
>the peaks of the signal, and sometimes it would coincide with
>the zero crossing points. How is this corrected later?
This "beat" represents information that is OUTSIDE the Nyquist
bandwidth. Look at the spectrum of your original signal with the
"beats" as you describe: in order for such beats to occur, there
must be components ABOVE the Nyquits frequency.
That's why, to complete the process, not only must you remove
components above the Nyquist frequency in the A/D process, you
MUST also remove them AFTER the D/A process. SO you have TWO
filters limiting the audio to below the Nyquist point: on on
A/D, the other on D/A. When that second filter is applied as it
MUST be, the neats go away because the components that make the
beats are eliminated by the filter.
>So when it is said that the input must be bandwidth limited to
>frequencies *less* than the Nyquist frequency, how much less is
>*less*? To me, it seems that the nearer an input frequency
>approaches the Nyquist frequency, the more error is possible.
>But this goes against many assertions I have read about
>sampling, so what fundamental am I missing?
You're missing the requirement for a Nyquist-freuqnecy filter at
TWO places, on A/D and on D/A.
Given that a properly dithered signal completely removes all
quantisation distortion, the advantage of longer word converters
is then potentially lower noise floor, greater dynamic range,
and *not* greater resolution? (Plenty of potential for error in
that sentence!)
Greater dynamic range IS greater resolution.
>I've read that full scale on a 16 bit converter is
>conventionally 2V. Is that right?
That's true of some converters, nut only by convention.
>What is full scale on a 20 bit or 24 bit converter?
It could be exactly the same. It could be comething different.
It all depends upon what the designer did. There is no
enforceable "standard."
>Have we divided the scale more
>finely, or enlarged the scale, or both?
It depends upon how the designer has chosen to normalize the
scale. If the design is such that the output is the same 2 V at
digital full scale, then in order to achieve 24 bit resolution,
the noise floor MUST be lower, lower by 48 dB or so. If the
noise floor is the same, then the output voltage at full scale
MUST be greater.
That lower noise represents the smallest unambiguous change in
voltage that represents real signal. Any signal lower than that
can't be distinguished from noise, all other things being
equal.
>CD distortion at low levels: What's the story? IIRC Noel
>Keywood(?) tests equipment for one of the UK audiophool
>magazines. He routinely reports that CD players have 30%
>THD(?) at -60dB or similar (I may well have remembered that
>incorrectly). How's he making this measurement?
Incorrectly. He's using THD, a fairly broad and not very useful
measurement in a way that makes the numbers completely and
absurdly useless.
>What does it mean?
It means he has no idea what THD means. THD measurements are
made by simply eliminating the fundamental of a sine wave tone
and measuring EVERYTHING that's left over: that means real
distortion, random noise, power supply hum, radio interference,
EVERYTHING. He is NOT measuring harmonic distortion, he is
measuring harmonic distortion, noise and all other
interferences all mixed together in a way you can't tell them
apart.
Not only that, but if his signal is not properly dithered (and
dithering is done as part of the RECORDING process, it is NOT a
function of CD playeres or D/A converters), he's meanuring all
the errors inherent in his process.
Further, try measuring the "THD" of an LP system at -60 dB: it
will be, under these same conditions, MUCH worse. And this
points aout the ansurdity of his "technique."
>I can't escape the feeling I'm overlooking things very basic &
>embarrassing.
Well, you'd hardly be the first not to get this stuff. It is NOT
intuitive at all, and depending upon intuition WILL take you
down a path that leads to incorrect and contradictory
answer. You're fortunate because you KNOW you might be getting
it wrong. Others, including quite a few magazine writers, think
they got it right, and go on to make these grand pronouncements
about how this stuff works that's horrible and embarassingly
wrong: and then poor readers such as yourself get further
confused.
Keep asking questions, then.
--
| Dick Pierce |
| Professional Audio Development |
| 1-781/826-4953 Voice and FAX |
| DPi...@world.std.com |
Johnny
<donot...@interbulletin.bogus> wrote:
>Hi,
>As per subject, I'm struggling with several aspects of digital audio theory, and I think you will able to illuminate me. Please jump in and correct me at any time, because I must be going wrong somewhere, and I want to know.
>
>I think of part of Nyquist as saying "You must have *more* than two samples per cycle of the highest frequency you want to record". (This is easier for my brain, but will thinking of it like this trap me later on?)
>So ideally, there would not be any information at the Nyquist frequency or above in the input. OK fine, no problem understanding that.
>
>Now imagine that the input filter had intentionally been left out, and a signal of exactly the Nyquist frequency was deliberately injected. This signal would not be recorded properly. It is possible that the sampling window would coincide with the zero crossing points of the signal, or any other point on the signal, extremely unlikely to be the -ve & +ve peaks, right?
>
>Then remove the first signal, and inject a signal of a frequency very close to, but just slightly less than the Nyquist frequency. This signal would then be being sampled at a rate slightly higher than 2 samples/cycle.
>Here's my problem - I think that the recording of this signal would beat. Sometimes the sampling window would coincide with the peaks of the signal, and sometimes it would coincide with the zero crossing points. How is this corrected later?
>
>So when it is said that the input must be bandwidth limited to frequencies *less* than the Nyquist frequency, how much less is *less*? To me, it seems that the nearer an input frequency approaches the Nyquist frequency, the more error is possible. But this goes against many assertions I have read about sampling, so what fundamental am I missing?
In theory it can approach the nyquist frequeny as closely as you like.
In practice, the *frequency response* is limitated by the
Anti-aliasing and reconstruction filters that can be used. As others
have explained this "beating" that you mention can be avoided by using
an antialiasing filter on the A->D recording process and a
Reconstruction filter on the D->A process.
Although the frequency response can potentially reach right up to the
nyquist frequency, the limited bandwidth prevents transient
information from being recorded at these frequencies. It is a basic
property of discrete time signal processing, that a bandwidth limited
system will have an infinately long reponse in the time domain. When
you deal with frequencies very close the Bandwidth limit of the system
the Time-domain performance becomes worse and worse. As the frequency
approaches the nyquist frequency, sound can only be recorded as a
single frequency that is continuous in the time domain - ie never
starts and never stops, which is lousy transient performance.
So basically how close you want to go to the nyquist frequency depends
on the accuracy of the transient response that you want to record.
Hope this helps,
Johnny.
Art Ludwig
Johnny <john...@one.net.au.NOSPAM> wrote in message
news:3abf0156...@news.one.net.au...
> On Sun, 25 Mar 2001 01:28:17 +0000, Clueless
> <donot...@interbulletin.bogus> wrote:
>
> >
> >Then remove the first signal, and inject a signal of a frequency very
close to, but just slightly less than the Nyquist frequency. This signal
would then be being sampled at a rate slightly higher than 2 samples/cycle.
> >Here's my problem - I think that the recording of this signal would beat.
Sometimes the sampling window would coincide with the peaks of the signal,
and sometimes it would coincide with the zero crossing points. How is this
corrected later?
> >
>
> In practice, the *frequency response* is limitated by the
> Anti-aliasing and reconstruction filters that can be used. As others
> have explained this "beating" that you mention can be avoided by using
> an antialiasing filter on the A->D recording process and a
> Reconstruction filter on the D->A process.
>
out
http://www.silcom.com/~aludwig/images/Sampling.gif
It illustrates a signal sampled close to the Nyquist frequency, with the
sample values beating as you note, and with the sinc functions which
reconstruct the signal. The reconstruction process is discussed further in
http://www.silcom.com/~aludwig/Signal_processing/Signal_processing.htm
Regards, Art Ludwig
Clueless
A few more basic questions if I may:
A long word converter theoretically allows a fantastically low noise floor? In reality, how much is realisable, given that associated analogue components probably cannot match it? I.e. how quiet can the best analogue gear be? (I'm not worried about noise, but I'm interested.)
If not all of the resolution is usable, what are the advantages of that extra resolution? Better for calculations in signal processing?
Related: 24/96 audio. As far as I can work out now, the main advantage of this (along with potentially lower noise) is that the requirements for the filters are far less severe, as they can be moved well away from the audible range. Sound right?
(In particular the claims of the audiophool magazines that the "graininess" of CD will be gone are rubbish?)
How steep can a practical input filter be made? In a real system where would the 3dB point be? In practise would a tiny amount of signal with frequency higher the Nyquist frequency be allowed into the sample & hold circuit?
Say I extract the audio from a CD to my PC with one of these DAE programs and look at it in a wave editor. Because this data hasn't gone through the reconstruction filter, what I'm looking at actually isn't representative of the final waveform, is it? (I'd think it fairly resembles the final waveform though?)
Noel Keywood (if that is his name) tests the frequency response of CD players with a test disc that contains an impulse. He deconvolves the impulse to get the frequency reponse. Is this a valid test? I notice sometimes the CD players are unable to play the impulse.
(His speaker frequency plots are 3rd octave, but I don't know how they are made. The impedance plots are continuous.)
Are there any publications relating to audio which have a reasonable technical foundation?
Thanks again for your help.
Richard D Pierce
Clueless <donot...@interbulletin.bogus> wrote:
>Thank you all for your informative replies. (Very high s/n
>ratio - what on earth are you doing on Usenet?!?)
Trying to lower the noise floor in the face of overwhelming
hooey!
>A long word converter theoretically allows a fantastically low
>noise floor? In reality, how much is realisable, given that
>associated analogue components probably cannot match it? I.e.
>how quiet can the best analogue gear be? (I'm not worried
>about noise, but I'm interested.)
Practically, let's see. Very high quality studio microphones
exist with a noise floor equivalent to about 20-25 dB SPL
unweighted as about the absolute minimum. SO even if we
calibrate the microphone and A/D system so that the least
significant bit is toggling at 0 dB SPL,. that means the bottom
3 bits of a theoretically perfect 24 bit converter is doing
nothing but measuring microphone noise, and we can still hit 144
dB SPL, a preposterously loud level.
>If not all of the resolution is usable, what are the advantages
>of that extra resolution? Better for calculations in signal
>processing?
You don't need the extra precisioon at recording time to take
advantage of a wider word size at processing time. I was lead
software engineer on a digital audio editing workstation that
had 16 bit converters that did its internal processing in 56 bit
precision.
>Related: 24/96 audio. As far as I can work out now, the main
>advantage of this (along with potentially lower noise) is that
>the requirements for the filters are far less severe, as they
>can be moved well away from the audible range. Sound right?
Partially. But that is also ROUTINELY done with 44.1 kHz audio
as well. The technique is called "oversampling." What either a
high sampling rate or oversampling buys you is that it moves the
aliases and images ABOVE the Nyquist frequency, but now that
Nyquist frequency is much higher. For example, a 64x oversampled
44.1 kHz system has a Nyquist frequency not at 22.05 kHz, but 64
times higher than that, at 1411.2 kHz. The analog portion of the
filter needed to deal withat is little more than a simple,
single pole, transient-perfect RC filter. However, the
requirement is STILL there it have a brick-wall filter at 22.05,
but now it can be done ENTIRELY in the digital domain, where you
have many more degrees of freedom at your disposal.
>(In particular the claims of the audiophool magazines that the
>"graininess" of CD will be gone are rubbish?)
The "graininess" has yet to be demonstrated as being an inherent
property of CDs. I have LPs that sound grainy. But it is NOT due
to the nonsense that these ill-informed writers claim it is.
>How steep can a practical input filter be made? In a real
>system where would the 3dB point be? In practise would a tiny
>amount of signal with frequency higher the Nyquist frequency be
>allowed into the sample & hold circuit?
When doing things in oversampled systems, the filters can be
made extremely sharp. The passband ripple can be a tiny fraction
of a dB, with a 3 dB down point of 21.5 kHz and the rejection
above the Nyquist frequency exceeding 90 dB.
>Say I extract the audio from a CD to my PC with one of these
>DAE programs and look at it in a wave editor. Because this
>data hasn't gone through the reconstruction filter, what I'm
>looking at actually isn't representative of the final waveform,
>is it? (I'd think it fairly resembles the final waveform
>though?)
Depends upon the editor. CoolEdit makes a reasonable attempt to
curve fit the data properly. But many editors don't even try and
the results are just plain wrong.
>Noel Keywood (if that is his name) tests the frequency response
>of CD players with a test disc that contains an impulse. He
>deconvolves the impulse to get the frequency reponse.
No, he doesn't. He takes the FFT to get complex transfer
function, then takes the magnitude of the complex trabsfer
function to get the frequency response.
>Is this a valid test?
Within limits. If the impulse on the CD is generated by a simple
computer program that runs a bunch of samples at 0, then a few
at some amplitude, and then a bunch more at 0, it is quite an
invalid test. This is because you have generated an impulse in
the digital domain that has components outside the Nyquist
frequency and thus MUST have aliasing built into it. In a
similar fashion, a computer generated square wave has precisely
the same problem.
>(His speaker frequency plots are 3rd octave, but I don't know
>how they are made. The impedance plots are continuous.)
There are a variety of systems that can do this.
Stewart Pinkerton
>Thank you all for your informative replies. (Very high s/n ratio - what on earth are you doing on Usenet?!?) Yes, as you've shown me, I didn't appreciate what the reconstruction filter was doing.
>
>
>A few more basic questions if I may:
>
>A long word converter theoretically allows a fantastically low noise floor? In reality, how much is realisable, given that associated analogue components probably cannot match it? I.e. how quiet can the best analogue gear be? (I'm not worried about noise, but I'm interested.)
Well, 24 bits is a 144dB range, so let's look at how you might
approach it. In purely electrical terms, you're not going to get much
below 1uV in a 20kHz bandwidth - that's just the thermal noise of a
3k3 resistor! Track this up by 144dB and you get around 15 volts, so
you'd need to run at a 0dB reference of say 20 volts rms to have any
eal hope of achieving the full dynamic range. Acoustically however,
you have a much more basic problem. Even the finest studio microphones
have a noise floor around 20dB absolute minimum, and no way will you
ever hear higher than 130dB at any musical event (more than 110dB is
*extremely* rare for non-amplified concerts), so you'll never actually
need more than 110dB to encompass the widest possible dynamic range in
the real world. Many leading authorities such as Bob Stuart have
suggested that 20 bits are all that will ever be required in the
playback medium, with 24 in the recording studio giving a comfortable
buffer for unexpected mic overloads and EQ.
Arny Krueger
"Clueless" <donot...@interbulletin.bogus> wrote in message
> A long word converter theoretically allows a fantastically low
noise floor?
In theory yes. In practice they produce lots of bits of noise.
> In reality, how much is realizable, given that associated analogue
components probably cannot match it?
The best converters claim about 120 dB dynamic range unweighted.
> I.e. how quiet can the best analogue gear be? (I'm not worried
about noise, but I'm interested.)
People talk about 132 dB, but back in the real world the really good
analog high-level signal processing gear has unweighted dynamic range
in the 95-105 dB range (referencing the standard level of +4 dB). It
is not unusual for gear to be able to handle levels 10-15 dB over +4,
so the total dynamic range of the best analog gear is observably in
the 105-120 dB range.
The weakest link is probably microphone preamps, some of which have
tremendous dynamic range if you allow the gain control to be adjusted
in the midst of measuring them, but in practice they have dynamic
range in the same range or a little less than I just mentioned,
which is really pretty impressive if you consider the size of the
voltages they handle at their inputs.
Mixing consoles with large numbers of inputs that can be combined do
far worse, and can often be found operating with dynamic range in the
75-90 dB range.
Home audio equipment typically has dynamic range in the 75-100 dB
range. The lower number would be tubed equipment, and the higher
number would be the better solid state stuff.
However this is all in the world of electronics. When you move into
real world recording studios and performance halls, the background
noise due to air conditinioing equipment, nearby traffic, and the
breathing and motion of people gets things down into the 70-80 dB
range pretty quickly.
> If not all of the resolution is usable, what are the advantages of
that extra resolution?
Not a whole bunch. In the end we do a nice job of transporting the
residual noise of the recording studio and mixing console into
people's living rooms.
> Better for calculations in signal processing?
Well, yes 24 bit or 32 bit arithmetic allows us to do lots of
shucking and jiving without harming the dynamic range of what was
recorded in the studio, for better or worse.
> Related: 24/96 audio. As far as I can work out now, the main
advantage of this (along with potentially lower noise) is that the
requirements for the filters are far less severe, as they can be
moved well away from the audible range. Sound right?
That's the theory. However the filter problem was pretty well solved
about a decade or more ago when ADCs and DACs switched over to using
digital filters. There is a theory that "brick wall" filtering does
nasty audible things at lower frequencies. You can evaluate that
theory for yourself by downloading and listening to files from
http://www.pcabx.com/technical/sample_rates/index.htm and/or
http://www.pcabx.com/technical/low_pass/index.htm . That theory is
not well-supported by the audible evidence.
> (In particular the claims of the audiophool magazines that the
"graininess" of CD will be gone are rubbish?)
That can only be true if 16/44 processing has inherent "graininess".
If the references I just gave don't convince you of that, there is
always http://www.cdabx.com/technical/bits44/index.htm which shows
the audible effects of audio coding with various numbers of bits and
various forms of bit truncation.
> How steep can a practical input filter be made?
Unbelievably steep. 100's of dB per octave. (digital filtering,
natch).
>In a real system where would the 3dB point be? In practice would a
tiny amount of signal with frequency higher the Nyquist frequency be
allowed into the sample & hold circuit?
In practice this is done and some small amounts of aliasing can be
found.
> Say I extract the audio from a CD to my PC with one of these DAE
programs and look at it in a wave editor. Because this data hasn't
gone through the reconstruction filter, what I'm looking at actually
isn't representative of the final waveform, is it? (I'd think it
fairly resembles the final waveform though?)
The wave editor I use, CoolEdit simulates the effect of the usual
filtering before it writes a wave to the screen. This is pretty
commonly done.
> Noel Keywood (if that is his name) tests the frequency response of
CD players with a test disc that contains an impulse.
This can be done.
>He deconvolves the impulse to get the frequency response. Is this a
valid test?
Yes.
>I notice sometimes the CD players are unable to play the impulse.
Yet another thing that Mr. Keywood sees that I've never seen. I've
seen some incredibly cheap digital hardware at least pass impulses.
The results weren't necessarily pretty, but the impulses were passed.
> (His speaker frequency plots are 3rd octave, but I don't know how
they are made. The impedance plots are continuous.)
Third octave is considered to be somewhat coarse by modern standards.
Continuous impedance plots are a piece of cake.
Clueless
>I wrote:
>>Related: 24/96 audio.
>>... the requirements for
>>the filters are far less severe, as they
>
>Partially. But that is also ROUTINELY done with 44.1 kHz audio
>as well. The technique is called "oversampling."
So at the listener's end, it's pretty well sorted then?
Just to clarify (I'm unsure), for 44.1kHz, you still need a steep analogue filter at the *recording* end of things? Or are they sampling at a much higher rate than 44.1kHz, filtering digitally, then knocking it back to 44.1k?
What's the steepest practical analogue filter?
>If the impulse on the CD is generated by a simple
>computer program that runs a bunch of samples at 0,
>then a few at some amplitude, and then a bunch more
>at 0, it is quite an invalid test.
If one were to "measure" frequency response in this way, what kind of result would one get? Something just totally off the wall, or something seemingly reasonable but wrong?
Cheers,
Clueless
>...the filter problem was pretty
>well solved about a decade or more ago when ADCs
>and DACs switched over to using digital filters.
The initial filter on an ADC has to be analogue, so does that mean they are sampling at a much higher rate than 44.1kHz, then filtering digitally? (I think I asked this in the other post.)
>There is a theory that "brick wall" filtering does
>nasty audible things at lower frequencies. You can evaluate that
>theory for yourself by downloading and listening to files from
>http://www.pcabx.com/technical/sample_rates/index.htm and/or
>http://www.pcabx.com/technical/low_pass/index.htm . That theory is
>not well-supported by the audible evidence.
..
>http://www.cdabx.com/technical/bits44/index.htm which
>shows the audible effects of audio coding with various
>numbers of bits and various forms of bit truncation.
I want to listen to these, but we don't have soundcards at school. (Even if we did, it's pretty noisy.) Thanks for the explanations of that and the other stuff though.
Cheers,
Geoff Wood
oversampling is employed, the analogue filter can be of a lower-order (with
better phase linearity) than the otherwise required 'brickwall' filter.
geoff
"Clueless" <donot...@interbulletin.bogus> wrote in message
news:3AC2A320...@interbulletin.com...
Stewart Pinkerton
>DPi...@world.std.com (Richard D Pierce) wrote:
>>I wrote:
>>>Related: 24/96 audio.
>>>... the requirements for
>>>the filters are far less severe, as they
>>>can be moved well away from the audible range...
>>
>>Partially. But that is also ROUTINELY done with 44.1 kHz audio
>>as well. The technique is called "oversampling."
>
>So at the listener's end, it's pretty well sorted then?
Pretty well - see the dCS 'RingDAC' for the current state of the art.
>Just to clarify (I'm unsure), for 44.1kHz, you still need a steep analogue filter at the *recording* end of things? Or are they sampling at a much higher rate than 44.1kHz, filtering digitally, then knocking it back to 44.1k?
The main advantage of oversampling is that the always necessary 'brick
wall' filter can be implemented in the digital domain, which is much
more repeatable in production.
>What's the steepest practical analogue filter?
The original CD equipment used up to twelve pole analogue filters,
were very expensive to make and tended to drift in production. Any
more than a five-pole (30dB/octave) filter gets seriously difficult to
make in mass production due to component tolerances and thermal drift.
>>If the impulse on the CD is generated by a simple
>>computer program that runs a bunch of samples at 0,
>>then a few at some amplitude, and then a bunch more
>>at 0, it is quite an invalid test.
>
>If one were to "measure" frequency response in this way, what kind of result would one get? Something just totally off the wall, or something seemingly reasonable but wrong?
Something with aliasing, since that unnaturally fast rise time cannot
occur on a real CD, due to the <22.05kHz upper frequency limit.
Arny Krueger
"Clueless" <donot...@interbulletin.bogus> wrote in message
> DPi...@world.std.com (Richard D Pierce) wrote:
> >I wrote:
> >>Related: 24/96 audio.
> >>... the requirements for
> >>the filters are far less severe, as they
> >>can be moved well away from the audible range...
> >Partially. But that is also ROUTINELY done with 44.1 kHz audio
> >as well. The technique is called "oversampling."
> So at the listener's end, it's pretty well sorted then?
YMMV. After all nothing but the free enterprise system keeps people
from turning out crap. The good news is that it can be well-sorted,
and for not unreasonable prices
> Just to clarify (I'm unsure), for 44.1kHz, you still need a steep
analogue filter at the *recording* end of things?
Yes, by some means.
> Or are they sampling at a much higher rate than 44.1kHz, filtering
digitally, then knocking it back to 44.1k?
Yes, the general run of ADCs are based on oversampling, too.
> What's the steepest practical analogue filter?
I seem to recall that the analog filter in the original CDP-101 was
printed up in the service manual but with no parts values. It was an
encapsulated unit. I seem to recall 12 coils and 12 capacitors so
that would be 24 poles (if memory serves). I'm not sure this was the
steepest practical analog filter because I don't know if it was the
most complex, and I don't know if it was that practical. You can see
how it worked at
http://www.pcavtech.com/play-rec/Sony_CDP-101/index.htm#FR_DA . By
modern standards (see
http://www.pcavtech.com/play-rec/CDP-XE_500/index.htm#FR_DA and
notice the vastly expanded vertical scale) it was not that practical.
> >If the impulse on the CD is generated by a simple
> >computer program that runs a bunch of samples at 0,
> >then a few at some amplitude, and then a bunch more
> >at 0, it is quite an invalid test.
> If one were to "measure" frequency response in this way, what kind
of result would one get?
Actually, I think you'd get something pretty close to the right
answer, but it might be a little off because it would show arpature
effects that are usually partially compensated by a real-world ADC.
Richard D Pierce
Clueless <donot...@interbulletin.bogus> wrote:
>DPi...@world.std.com (Richard D Pierce) wrote:
>>I wrote:
>>>Related: 24/96 audio.
>>>... the requirements for
>>>the filters are far less severe, as they
>>>can be moved well away from the audible range...
>>
>>Partially. But that is also ROUTINELY done with 44.1 kHz audio
>>as well. The technique is called "oversampling."
>
>So at the listener's end, it's pretty well sorted then?
>
>Just to clarify (I'm unsure), for 44.1kHz, you still need a
>steep analogue filter at the *recording* end of things? Or are
>they sampling at a much higher rate than 44.1kHz, filtering
>digitally, then knocking it back to 44.1k?
Nope, oversampling works just fine at the A/D end of things as
well.
>What's the steepest practical analogue filter?
Define "practical."
>>If the impulse on the CD is generated by a simple
>>computer program that runs a bunch of samples at 0,
>>then a few at some amplitude, and then a bunch more
>>at 0, it is quite an invalid test.
>If one were to "measure" frequency response in this way, what
>kind of result would one get? Something just totally off the
>wall, or something seemingly reasonable but wrong?
Difficult to say: all of the out-of-band artifacts would be
folded back into the audi band, but the spectrum of an inpulse
is dense abyway, so , in a raw frequency response, you might not
see anything.
Clueless
>There is always an analogue filter on the input to the ADC,
>however if oversampling is employed, the analogue filter
>can be of a lower-order (with better phase linearity) than
>the otherwise required 'brickwall' filter.
Thank you all for your patience in answering what must be to you trivial questions. I think I'm starting to get a handle on the basics now.
Now I'm guessing that since the phase shift of the analogue filter is a known quantity, it can be corrected in the digital domain at the same time as the digital filtering?
Well I'm going to go away and do some more reading, and try and attach some numbers to some of these concepts, then maybe I'll come bug you guys again. Thanks again,
Geoff Wood
phase anomalies, but can't cause 'ringing', evident with high amplitude HF
signals. This is one of the reasons the first generation of CD players
sounded 'harsh'.
geoff
"Clueless" <donot...@interbulletin.bogus> wrote in message
news:3AC51B5C...@interbulletin.com...
Jay - AtlDigi
<donot...@interbulletin.bogus> wrote:
> "Geoff Wood" wrote:
> >There is always an analogue filter on the input to the ADC,
> >however if oversampling is employed, the analogue filter
> >can be of a lower-order (with better phase linearity) than
> >the otherwise required 'brickwall' filter.
>
> Thank you all for your patience in answering what must be to you trivial
> questions. I think I'm starting to get a handle on the basics now.
>
> Now I'm guessing that since the phase shift of the analogue filter is a
> known quantity, it can be corrected in the digital domain at the same
> time as the digital filtering?
>
> Well I'm going to go away and do some more reading, and try and attach
> some numbers to some of these concepts, then maybe I'll come bug you guys
> again. Thanks again,
I have some articles at my website that deal with these issues if you
are interested in something additional to read. www.promastering.com
and click on "tech talk".
-Jay Frigoletto
http://www.promastering.com
Stewart Pinkerton
>One thing not mentioned before is not only that high order filters have
>phase anomalies, but can't cause 'ringing', evident with high amplitude HF
>signals. This is one of the reasons the first generation of CD players
>sounded 'harsh'.
To avoiding confusing our new friend, let's point out that you did of
course mean that high-order filters *can* cause ringing! :-)
>
>geoff
>
>"Clueless" <donot...@interbulletin.bogus> wrote in message
>news:3AC51B5C...@interbulletin.com...
>
>>
>> Thank you all for your patience in answering what must be to you trivial
>questions. I think I'm starting to get a handle on the basics now.
>>
>> Now I'm guessing that since the phase shift of the analogue filter is a
>known quantity, it can be corrected in the digital domain at the same time
>as the digital filtering?
Yes.
Geoff Wood
wot I reely meant.
g. ;-)
"Stewart Pinkerton" <pat...@popmail.dircon.co.uk> wrote in message
news:3ac57c7e....@news.freeserve.net...
Richard D Pierce
>>There is always an analogue filter on the input to the ADC,
>>however if oversampling is employed, the analogue filter
>>can be of a lower-order (with better phase linearity) than
>>the otherwise required 'brickwall' filter.
>
>Thank you all for your patience in answering what must be to
>you trivial questions. I think I'm starting to get a handle on
>the basics now.
You're quite welcome
>Now I'm guessing that since the phase shift of the analogue
>filter is a known quantity, it can be corrected in the digital
>domain at the same time as the digital filtering?
Or, more significantly, since the analog filter requirements are
so loose, there may not be any need to correct the phase shift
in the digital domain.