[Randy Yates]

Robert Orban writes:

In article ,
says...

There is no nonlinear distortion in a digital filter unless you screw up and
overdrive it. Digital filters generally have far more dynamic range than
analog filters, so it is unlikely that it will clip unless someone really
screws up.

Virtually all digital filters have some sort of nonlinear distortion. The IIRs
typically used in equalizers are subject to limit cycles and truncation noise
due to finite length arithmetic. There are various techniques (like error
feedback) for minimizing these issues, but it is very rare to find a digital
filter in the real world that in fact adds dither before every truncation,
which is what would be required to linearize the filter.

These problems all increase as the poles of the IIR filter get closer to the
unit circle in the z plane. This is a strong argument for using as low a
sample rate as you can get away with because doing so minimizes the noise in
any filters. In audio, this includes EQ as well as crossovers in multiband
compressors. In FIR filters, there is a similar advantage to minimizing the
sample rate because it reduces the length of the filter required to produce a
given target response.

If you are doing nonlinear processing, like compression and limiting, then you
can get aliasing in the sidechains and also when the audio is multiplied by
the sidechain signal. So here the advice is different -- use a sampling rate
that is high enough to render any such aliasing inaudible.

Hi Robert,

An excellent summary. I would add that the nonlinearities in IIRs can be
(and are typically) much greater than those in FIRs. The basic
nonlinearity in an FIR is requantization from a high word length (the
accumulated word length) back down to the data path width (which is
relatively benigh), whereas nonlinearities in IIRs can include, in
addition to those you stated (which are already significantly worse than
FIRs' requantization) saturation and "bit sticking," which are
particularly nasty unless extreme care is taken in the implementation.

Regarding sample rate, I agree wholeheartedly with your
conclusion. Stick with 44.1 kHz - not only is it not bad, it's better!

I've also wanted to chime in here and state that there is absolutely no
need to maintain an oversampled rate (e.g., 96 kHz) for the purpose of
relaxing the analog anti-aliasing and reconstruction filter
requirements; this can be accomplished at the front- and back-ends while
maintaining a mundane 44.1 or 48 kHz in the middle.
--
% Randy Yates % "The dreamer, the unwoken fool -
%% Fuquay-Varina, NC % in dreams, no pain will kiss the brow..."
%%% 919-577-9882 %
%%%% % 'Eldorado Overture', *Eldorado*, ELO
http://www.digitalsignallabs.com

[Laurence Payne[_2_]]

On Mon, 27 Oct 2008 23:26:29 GMT, Blind Hog
wrote:

My first action in the editing software is to expand to 24 bits. My
last action before burning a CD is to convert back to 16 bits.

Both of which are surely inevitable (and transparant to the operator)
in any decent multitrack editor? Except that the internal format is
more likely to be 32-bit.

[Blind Hog[_2_]]

"geoff" wrote in
:

What recording device are you using ? Obviously available media/memory
is a restriction with it for you.....

I use an Edirol R-4 with (I think) a 30 gb hard disk.

24 bits would allow less fretting over level-setting optimisation, and
obviate the necessity for extra processing should ypou feel the need
to convert.

But 16 is enough and if I'm more than 30 dB off level, I'm hearing a lot
of complaint from my preamps.

I can't imagine that you would be getting any benefit from converting
to 24 bits before any processing. But that depends, I guess, on how
primitive your editing app is. The processing 'bitdepth' does not
necessarily have to have a fixed relationship with source media file
bit-dept.

I use Audition.

I can hear a difference when I add effects like noise reduction and
reverb that trail to zero. I don't know of any software that will
perform a process (regardless of the bit depth) that can cleanly perform
to the last bit or two of resolution. Better to do that work in more
bits, perform all the operations, then dither to 16 and cover every math
error.

[Geoff]

Blind Hog wrote:
I can hear a difference when I add effects like noise reduction and
reverb that trail to zero. I don't know of any software that will
perform a process (regardless of the bit depth) that can cleanly
perform to the last bit or two of resolution. Better to do that work
in more bits, perform all the operations, then dither to 16 and cover
every math error.

Well there are reverb plugins and reverb plugins, and not all are the same,
or use the same principles. Some may be vastly superior (or not) to what you
are using, and may (or may not) do those last 2 bits that you are worried
about.

And by the way, changing from 16 to 24 bits is NOT just adding extra zeros
at the bottom end.

geoff

[Scott Dorsey]

Robert Orban wrote:
In article ,
says...

There is no nonlinear distortion in a digital filter unless you screw up and
overdrive it. Digital filters generally have far more dynamic range than
analog filters, so it is unlikely that it will clip unless someone really
screws up.

Virtually all digital filters have some sort of nonlinear distortion. The IIRs
typically used in equalizers are subject to limit cycles and truncation noise
due to finite length arithmetic. There are various techniques (like error
feedback) for minimizing these issues, but it is very rare to find a digital
filter in the real world that in fact adds dither before every truncation,
which is what would be required to linearize the filter.

It's becoming increasingly popular to do this in DAW software, though, and
the increasing move to floating point implementations on DAW systems change
the precision issues a little bit (although they don't eliminate them).

These problems all increase as the poles of the IIR filter get closer to the
unit circle in the z plane. This is a strong argument for using as low a
sample rate as you can get away with because doing so minimizes the noise in
any filters. In audio, this includes EQ as well as crossovers in multiband
compressors. In FIR filters, there is a similar advantage to minimizing the
sample rate because it reduces the length of the filter required to produce a
given target response.

If you are doing nonlinear processing, like compression and limiting, then you
can get aliasing in the sidechains and also when the audio is multiplied by
the sidechain signal. So here the advice is different -- use a sampling rate
that is high enough to render any such aliasing inaudible.

These are both extremely true and they are also both good arguments for doing
the actual processing with floating point data.
--scott
--
"C'est un Nagra. C'est suisse, et tres, tres precis."

[Randy Yates]

(Scott Dorsey) writes:
[...]
Robert Orban wrote:
If you are doing nonlinear processing, like compression and limiting, then you
can get aliasing in the sidechains and also when the audio is multiplied by
the sidechain signal. So here the advice is different -- use a sampling rate
that is high enough to render any such aliasing inaudible.

These are both extremely true and they are also both good arguments for doing
the actual processing with floating point data.

Huh? Even infinite-precision arithmetic won't help with these. The
problem is one of aliasing, not resolution.
--
% Randy Yates % "Rollin' and riding and slippin' and
%% Fuquay-Varina, NC % sliding, it's magic."
%%% 919-577-9882 %
%%%% % 'Living' Thing', *A New World Record*, ELO
http://www.digitalsignallabs.com

[Scott Dorsey]

Randy Yates wrote:
(Scott Dorsey) writes:
[...]
Robert Orban wrote:
If you are doing nonlinear processing, like compression and limiting, then you
can get aliasing in the sidechains and also when the audio is multiplied by
the sidechain signal. So here the advice is different -- use a sampling rate
that is high enough to render any such aliasing inaudible.

These are both extremely true and they are also both good arguments for doing
the actual processing with floating point data.

Huh? Even infinite-precision arithmetic won't help with these. The
problem is one of aliasing, not resolution.

Ahh, seperate issue, then! I thought he was referring to a mathematical
precision issue causing aliasing, not the harmonic distortion inherent in
any nonlinear system.

In the analogue world, though, a lot of people _like_ the distortion products
caused by compression and limiting.
--scott


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

[Randy Yates]

(Scott Dorsey) writes:

Randy Yates wrote:
(Scott Dorsey) writes:
[...]
Robert Orban wrote:
If you are doing nonlinear processing, like compression and limiting, then you
can get aliasing in the sidechains and also when the audio is multiplied by
the sidechain signal. So here the advice is different -- use a sampling rate
that is high enough to render any such aliasing inaudible.

These are both extremely true and they are also both good arguments for doing
the actual processing with floating point data.

Huh? Even infinite-precision arithmetic won't help with these. The
problem is one of aliasing, not resolution.

Ahh, seperate issue, then! I thought he was referring to a mathematical
precision issue causing aliasing, not the harmonic distortion inherent in
any nonlinear system.

In the analogue world, though, a lot of people _like_ the distortion products
caused by compression and limiting.

Yes, but it is not those distortion products that Robert was attempting
to mitigate.

If you want to implement compression or limiting in the digital domain,
it helps to first upsample the signal before applying the nonlinearity,
then decimate back to the original sample rate afterwards. In this way,
fewer of the harmonics that result from the nonlinearity are aliased
back.
--
% Randy Yates % "She has an IQ of 1001, she has a jumpsuit
%% Fuquay-Varina, NC % on, and she's also a telephone."
%%% 919-577-9882 %
%%%% % 'Yours Truly, 2095', *Time*, ELO
http://www.digitalsignallabs.com

[Blind Hog[_2_]]

Laurence Payne wrote in
:

On Mon, 27 Oct 2008 23:26:29 GMT, Blind Hog
wrote:

My first action in the editing software is to expand to 24 bits. My
last action before burning a CD is to convert back to 16 bits.

Both of which are surely inevitable (and transparant to the operator)
in any decent multitrack editor? Except that the internal format is
more likely to be 32-bit.

Inevitable?

If I'm working in 16 bits the data is in 16 bits. Surely it upsamples to
perform and operation and downsamples again afterward, but I don't want the
downsample to happen but once--just before I burn the CD.

[Blind Hog[_2_]]

"geoff" wrote in
:

Blind Hog wrote:
I can hear a difference when I add effects like noise reduction and
reverb that trail to zero. I don't know of any software that will
perform a process (regardless of the bit depth) that can cleanly
perform to the last bit or two of resolution. Better to do that work
in more bits, perform all the operations, then dither to 16 and cover
every math error.

Well there are reverb plugins and reverb plugins, and not all are the
same, or use the same principles. Some may be vastly superior (or not)
to what you are using, and may (or may not) do those last 2 bits that
you are worried about.

I'm not worried about them because the bottom 8 bits disappear entirely
just before I downsample to 16 bits.

And by the way, changing from 16 to 24 bits is NOT just adding extra
zeros at the bottom end.

In what way is it not?

[Laurence Payne[_2_]]

On Wed, 29 Oct 2008 23:08:23 GMT, Blind Hog
wrote:

My first action in the editing software is to expand to 24 bits. My
last action before burning a CD is to convert back to 16 bits.

Both of which are surely inevitable (and transparant to the operator)
in any decent multitrack editor? Except that the internal format is
more likely to be 32-bit.

Inevitable?

Yup. They mix in 32-bits - you don't get any choice about it. A
16-bit input isn't "expanded" exactly, it's just put in an environment
where there's LOADS of headroom. The only place you have to worry
about keeping under nominal zero is the output bus.

But then you aren't expanding either - just padding out with 4 more
empty bits. If your editor is one that mixes in a high bit-depth
environment there's no point in you bothering.

[Les Cargill]

Blind Hog wrote:
"geoff" wrote in
:

Blind Hog wrote:
I can hear a difference when I add effects like noise reduction and
reverb that trail to zero. I don't know of any software that will
perform a process (regardless of the bit depth) that can cleanly
perform to the last bit or two of resolution. Better to do that work
in more bits, perform all the operations, then dither to 16 and cover
every math error.
Well there are reverb plugins and reverb plugins, and not all are the
same, or use the same principles. Some may be vastly superior (or not)
to what you are using, and may (or may not) do those last 2 bits that
you are worried about.

I'm not worried about them because the bottom 8 bits disappear entirely
just before I downsample to 16 bits.


Save 'em off seperately. They're nothing but noise....

And by the way, changing from 16 to 24 bits is NOT just adding extra
zeros at the bottom end.

In what way is it not?

--
Les Cargill