[Mogens V.]

wrote:
On Oct 30, 12:53 pm, Mark wrote:

What it buys you is that you no longer need a brick wall filter at 20 KHz...
you can instead have a very relaxed filter at a few hundred KHz if not
a couple MHz, and do most of the anti-aliasing filtration in the digital
domain when you decimate down from the high rate data down to 44.1.

A brick wall filter in the digital domain is just as bad as a brick
wall filter in the digital domain. Even an FIR brick wall filter will
ring... Gibbs phenomenon.

But all this ringing is up around 20 kHz and only happens when the
rest of the system is flat, so in practice sampling at 44.1kHz is just
fine.

Mark


and it is this artifact of filter ring (processing algorithm?) that
colors the digital sound,
thus, to avoid coloring sound with digital artifacts who's creation is
inherent in digital processing,

This thread doesn't evidence digital artefacts as inherent..

working at higher sample rates removes the artifacts from our audible
range.
leaving the air surrounding the note intact.

Provided your chosen converter happens to work better at such higher
sample rates..


Following the filter discussions.. are filters generally designed with
a fixed cutoff, or do they move up with (user selectable) sample rates?
If fixed, I may see a point in using higher sample rates - on an ill
designed converter, that is..

--
Kind regards,
Mogens V.

[Scott Dorsey]

Mogens V. wrote:

A brick wall filter in the digital domain is just as bad as a brick
wall filter in the digital domain. Even an FIR brick wall filter will
ring... Gibbs phenomenon.

But all this ringing is up around 20 kHz and only happens when the
rest of the system is flat, so in practice sampling at 44.1kHz is just
fine.

If properly designed, we can throw the ringing up above the audible band
and turn it into a non-issue. And, of course, with a digital filter, we
can eliminate group delay in the passband (which used to be a big deal).

Following the filter discussions.. are filters generally designed with
a fixed cutoff, or do they move up with (user selectable) sample rates?
If fixed, I may see a point in using higher sample rates - on an ill
designed converter, that is..

On properly designed systems, they move up. But, in the age of oversampling
where the majority of filtering is digital, there are some converters that
use a fixed analogue pre-filter. This is fine since the analogue filter is
operating well above the audio passband anyway.

There are some older designs from the ladder DAC era that used the same
filters for 44.1 and 48 ksamp/sec operation, resulting in substantial
changes in sound with small changes in sample rate. This sort of design
was inexcuseably sloppy back in 1985 when Panasonic was doing it, and it
remains so today.
--scott

--
"C'est un Nagra. C'est suisse, et tres, tres precis."

[[email protected]]

On Oct 30, 2:22 pm, "Mogens V."
wrote:
wrote:
On Oct 30, 12:53 pm, Mark wrote:

What it buys you is that you no longer need a brick wall filter at 20 KHz...
you can instead have a very relaxed filter at a few hundred KHz if not
a couple MHz, and do most of the anti-aliasing filtration in the digital
domain when you decimate down from the high rate data down to 44.1.

A brick wall filter in the digital domain is just as bad as a brick
wall filter in the digital domain. Even an FIR brick wall filter will
ring... Gibbs phenomenon.

But all this ringing is up around 20 kHz and only happens when the
rest of the system is flat, so in practice sampling at 44.1kHz is just
fine.

Mark

and it is this artifact of filter ring (processing algorithm?) that
colors the digital sound,
thus, to avoid coloring sound with digital artifacts who's creation is
inherent in digital processing,

This thread doesn't evidence digital artefacts as inherent..

working at higher sample rates removes the artifacts from our audible
range.
leaving the air surrounding the note intact.

Provided your chosen converter happens to work better at such higher
sample rates..

Following the filter discussions.. are filters generally designed with
a fixed cutoff, or do they move up with (user selectable) sample rates?
If fixed, I may see a point in using higher sample rates - on an ill
designed converter, that is..

--
Kind regards,
Mogens V.

if you are multiplying numbers,0&1's, then rounding them off /
truncating / artifacts.

[Mogens V.]

Scott Dorsey wrote:
Mogens V. wrote:

A brick wall filter in the digital domain is just as bad as a brick
wall filter in the digital domain. Even an FIR brick wall filter will
ring... Gibbs phenomenon.

But all this ringing is up around 20 kHz and only happens when the
rest of the system is flat, so in practice sampling at 44.1kHz is just
fine.


If properly designed, we can throw the ringing up above the audible band
and turn it into a non-issue. And, of course, with a digital filter, we
can eliminate group delay in the passband (which used to be a big deal).

Yup, I remember my '84 digital delay examn project. Used an 11th order
Sallen-Key bessel filter to get 18Khz bandwidth from 50Khz sample rate.

--
Kind regards,
Mogens V.

[Don Pearce]

On Tue, 30 Oct 2007 10:05:39 -0700, Mark wrote:

On Oct 30, 8:46 am, (Don Pearce) wrote:
On Tue, 30 Oct 2007 06:39:59 -0700, Mark wrote:
On Oct 30, 8:22 am, Mike Rivers wrote:
On Oct 30, 9:03 am, Randy Yates wrote:

Well, the simplest explanation is that the front-end (low-resolution)
ADC in a delta-sigma data converter really does sample at M*F_s samples
per second. I guess the point you're missing is that this is a very
coarse sampler - usually only 1 bit.

Yes, I understand that. It doesn't record the whole value each sample,
only whether it's greater or less than the previous sample.

The problem is that delta sigma data converters are not simple; it is
not a one- or two-paragraph topic.

So in this case we should just accept that they work and not argue
about how they work or why. That's OK with me. But when someone
suggests that they are effectively sampling at a higher rate, I
wondered why they just don't call it a higher sample rate and be done
with it? And the answer (again, simplified) is that the samples don't
actually provide any extended frequency information. They only help
the converters obey the primary law of sampling. To me, that's not
really working at a higher sample rate, it's using some clever
manipulation (a trick, in the colloquial jargon) to do what needs to
be done in order to make sampling work the way it should.

One doesn't need to understand the mechanism of delta-sigma converters
to understand what's really important about the technique.

someone briefly mentioned the key point..."steep filtering"

Steep filtering will create bad transient responses.

This is true if the steep filtering is in the analog domain or in the
digital domain.

If you sample at 44.1kHz or decimate down to 44.1 kHz, you need to
filter out everything above about 22 kHz. This "steep filter" can
create poor transient response near 20kHz IF the rest of the system
is flat. If the overall system response is not steeply filtered due
to some gradual rolloffs elsewhere like in the speakers or mic or
anywhere OR YOUR EAR...then the overall system is not steeply
filtered and there will not be poor transient responses. If the
sampling rate is 96kHz, then you don't need the steep filter in the
first place.

In THEORY 96kHz sampling can have better transient response around 20
kHz due to the lack of a steep anti-alias filter.

Not even in theory, unless your theory is very poor.

In THEORY 44.1 kHz sampling may have poorer transient response at
~20kHz due to the steep anti-alias filter (either analog or digital).

Again, not even in theory. Actually I guess you could say that it
doesn't actually have a transient response at 20kHz.



In PRACTICE sampling at 44.1 kHz will not have poor transient response
at 20 kHz because there is some other gradual rolloff in the system
such that the overall system response is not steeply rolled off.

Mark

No, that isn't why. In practice you don't get poor transients because
the filters are designed to have good transients. The gradual rolloff
you speak of is the analogue rolloff, which only needs to prevent
aliasing in the oversampled domain - way outside the audio range, and
it has no effect within the final 0 - 20kHz range.

d


If you sample at 96 kHz and then decimate to 44.1, you still need the
brickwall 20 kHz low pass filter, it's just at a different place.

___ANY___ low pass filter that has a passband to 20kHz and stop band
at 22kHz will RING at around 20kHz when hit with an impulse or a step.

You can play with the phase response, add a phase equalizer, or use
FIR or IIR or whatever you want and all you will be able to do is move
the ringing around to occur before or after the impulse (as seen at
the output) but you cannot get RID of the ringing except by making the
transition band (roll off) more gradual.

Again see Gibbs phenoemenon. Steep roll off = ringing ALWAYS!

Please show me one counterexample.


Mark


Gibbs phenomenon isn't ringing.

d

--
Pearce Consulting
http://www.pearce.uk.com

[Mike Rivers]

On Oct 30, 10:28 am, (Scott Dorsey) wrote:

One-bit sigma-delta systems DO get all those samples,
but they are only one bit long. And the sample is representing the slope
and not the instantaneous value.

Since this seems to be the nitpicking thread of the week, the bits in
a signa-delta stream don't represent the slope, they represent the
direction of the slope. I don't know how the three states of the
sample - less than the previous one, greater than the previous one, or
less-different-than-I-can-resolve - can be represented by one bit
that's either 1 or 0, but I'm sure the converter is smart enough to
figure it out.

[Mark]

On Oct 30, 2:22 pm, (Don Pearce) wrote:
On Tue, 30 Oct 2007 10:05:39 -0700, Mark wrote:
On Oct 30, 8:46 am, (Don Pearce) wrote:
On Tue, 30 Oct 2007 06:39:59 -0700, Mark wrote:
On Oct 30, 8:22 am, Mike Rivers wrote:
On Oct 30, 9:03 am, Randy Yates wrote:

Well, the simplest explanation is that the front-end (low-resolution)
ADC in a delta-sigma data converter really does sample at M*F_s samples
per second. I guess the point you're missing is that this is a very
coarse sampler - usually only 1 bit.

Yes, I understand that. It doesn't record the whole value each sample,
only whether it's greater or less than the previous sample.

The problem is that delta sigma data converters are not simple; it is
not a one- or two-paragraph topic.

So in this case we should just accept that they work and not argue
about how they work or why. That's OK with me. But when someone
suggests that they are effectively sampling at a higher rate, I
wondered why they just don't call it a higher sample rate and be done
with it? And the answer (again, simplified) is that the samples don't
actually provide any extended frequency information. They only help
the converters obey the primary law of sampling. To me, that's not
really working at a higher sample rate, it's using some clever
manipulation (a trick, in the colloquial jargon) to do what needs to
be done in order to make sampling work the way it should.

One doesn't need to understand the mechanism of delta-sigma converters
to understand what's really important about the technique.

someone briefly mentioned the key point..."steep filtering"

Steep filtering will create bad transient responses.

This is true if the steep filtering is in the analog domain or in the
digital domain.

If you sample at 44.1kHz or decimate down to 44.1 kHz, you need to
filter out everything above about 22 kHz. This "steep filter" can
create poor transient response near 20kHz IF the rest of the system
is flat. If the overall system response is not steeply filtered due
to some gradual rolloffs elsewhere like in the speakers or mic or
anywhere OR YOUR EAR...then the overall system is not steeply
filtered and there will not be poor transient responses. If the
sampling rate is 96kHz, then you don't need the steep filter in the
first place.

In THEORY 96kHz sampling can have better transient response around 20
kHz due to the lack of a steep anti-alias filter.

Not even in theory, unless your theory is very poor.

In THEORY 44.1 kHz sampling may have poorer transient response at
~20kHz due to the steep anti-alias filter (either analog or digital).

Again, not even in theory. Actually I guess you could say that it
doesn't actually have a transient response at 20kHz.

In PRACTICE sampling at 44.1 kHz will not have poor transient response
at 20 kHz because there is some other gradual rolloff in the system
such that the overall system response is not steeply rolled off.

Mark

No, that isn't why. In practice you don't get poor transients because
the filters are designed to have good transients. The gradual rolloff
you speak of is the analogue rolloff, which only needs to prevent
aliasing in the oversampled domain - way outside the audio range, and
it has no effect within the final 0 - 20kHz range.

d

If you sample at 96 kHz and then decimate to 44.1, you still need the
brickwall 20 kHz low pass filter, it's just at a different place.

___ANY___ low pass filter that has a passband to 20kHz and stop band
at 22kHz will RING at around 20kHz when hit with an impulse or a step.

You can play with the phase response, add a phase equalizer, or use
FIR or IIR or whatever you want and all you will be able to do is move
the ringing around to occur before or after the impulse (as seen at
the output) but you cannot get RID of the ringing except by making the
transition band (roll off) more gradual.

Again see Gibbs phenoemenon. Steep roll off = ringing ALWAYS!

Please show me one counterexample.

Mark

Gibbs phenomenon isn't ringing.

d

--
Pearce Consultinghttp://www.pearce.uk.com- Hide quoted text -

- Show quoted text -

Yes Gibbs phenomenon is ringing... I'm sure you hav elooked it up,
what would you call it.

Not all ringing is Gibbs phenomenon,,

Gibbs phenomenon is one example of a kind of ringing.

Mark

[Tom[_13_]]

On Oct 30, 10:22 pm, Mike Rivers wrote:
But when someone
suggests that they are effectively sampling at a higher rate, I
wondered why they just don't call it a higher sample rate and be done
with it? And the answer (again, simplified) is that the samples don't
actually provide any extended frequency information. They only help
the converters obey the primary law of sampling.

A delta-sigma converter really is sampling at a higher rate. I think
that its action can be explained pretty simply.
The quantizer has a fairly small number of bits (sometimes 1, but
usually "a few" in today's good converters). This is because it is
very difficult for physical reasons to quantize both quickly (high
sample rate) and accurately (large number of correct bits). Anyway, as
a result of the small number of bits, the total quantization noise
energy is high. This quantization noise would normally have a flat
spectrum from DC up to half the oversampled sampling frequency.
Therefore the majority of this noise isn't in the audio band anyway.
However, the remaining noise that is in the audio band is still too
high, so noise shaping is used to change the spectrum of the
quantization noise from flat to a curve so that almost all of it ends
up outside the audio band. This noise shaping is done using a feedback
loop. This is the "greater or less than the previous sample" action
that you refer to, but you don't need to think of it in such a
fundamental way - it's just a way of shaping the quantisation noise.
Then, this out-of-band noise needs to be removed. In a DAC this might
be done with a switched-capacitor filter or external filter. In an ADC
this is done by a digital low-pass filter, usually followed by
decimation to whatever final sample rate you want to record the data
at.

So the delta-sigma really is recording the data at a high rate. It's
just that it has to do something with the quantisation noise, so it
chooses to stick it up at the high frequencies and then filter out the
entire upper band. As such it can't be used to capture very ultrasonic
frequencies. If you were to bypass the noise shaping (whether this is
possible in a converter chip that you can actually buy is another
matter), you could capture the entire band from DC to half the
oversampled sampling frequency, but you'd have to contend with a
fairly high (and flat) quantization noise floor.
To make out that the whole thing is a signal processing trick or some-
such is rather missing the point in my opinion.

Tom

[Randy Yates]

Tom writes:
[...]

NICE explanation, Tom! And I think you also identified and answered Mike's implicit
question(s) as simply as possible. Well done.
--
% Randy Yates % "She's sweet on Wagner-I think she'd die for Beethoven.
%% Fuquay-Varina, NC % She love the way Puccini lays down a tune, and
%%% 919-577-9882 % Verdi's always creepin' from her room."
%%%% % "Rockaria", *A New World Record*, ELO
http://www.digitalsignallabs.com