In general, with audio it is better to sample at high depth and resolution
and then downgrade afterwards than it is to stay true to the final output.

Jon

I followed all the postings, and there are a lot of fine comments out
there. I guess I'll add my 2 bits:

First, going back to some of the earlier dither comments: The concept
of digital dither was covered pretty well - word length reduction. I
would probably magnify the fact that reduction with no dither causes 2
bad things. The first is harmonic distortions, the second is noise
modulation. Think for a second of a tiny sine wave, less than 1
quantization peak to peak, residing between 2 quantization levels. It
yields nothing. That is and incorect result. Now shift it so it
"seats" on a quantization level, going over and under and over and
under... by only one bit. That gives you a 1 bit amolitude square
wave... This is not a good signal either... This is just a starting
point to think about. Again you can see pictures on
www.lavryengineering.com (under suport, "Do you need 20 Bits" artical.

But I was more interested in analog dither. Years ago, I was
experimenting with in. Some of it was around Nyquist frequencies, some
was "subtuctive dither" (remove much in the digital domain what you
injected in the analog domain). Why analog dither? Well, nothing
happens if you feed a signal into a circuit and it does not cvause
some digital transitions... Say your circuit generates a different
code every 1 millivolt, but the signal is small and is always between
5.1 and 5.7mV... No action. So you need to "force it" to go over and
under some threshold (transition) levels. So we add noise. It does get
pretty complicated in a hurry. First knee jerk reaction is to add all
the noise in high frequncey. It is not that simple, because after
truncation, a lot of the high frequncey reapears in the audio range.

But you "have to do somthing"- add some random noise, thus trade off
some noise (dynamic range). For word reduction it is done to fix the
distortions and noise modulation problems. On the analog side, it was
about "getting anything to move" (and more than that). So the question
beacme, was is the best tradeoff? Least energy added for fixing the
problem? That was when triangle dither (2 least significant bits peak
to peak) solution came about.

But than there was a period where analog dither died (at least for
me). The quantization levels got so close to each other, that the
circuit noise was way bigger than I wished for. Take say a 10V range
and try for 20 bits - 10uV steps
(10 microvolts). Getting the signal in there with such low noie (10 uV
peak to peak) is not a walk in the park... You "call" the noise "too
much dither" and move farward...

But than analog dither came back - BIG TIMES! It started with 1 bit
(like DSD) and is now an everyday thing wth the newer multibit ADC's.
Those have just very few bits, thus the quantization levels are very
far apart. No longer in the microvolts, but in the volts (or near it).
This are all noise shaping converters. You feed back a filtered
quantized output and subtuct from the input.. This will take too long
to explain, but such a raw structure will have all sorts of problems
thus it requires dither, and often very large amplitude. The dither
randomness tend to "break up" what would otherwise be consistant
repetative patterns (called limit cycles), not exactly what you want
in the audio...

Regarding the comments on dynamic range and bits. Yes it was well
said, and about 6dB per bit is right. So 24 bits is about 144dB, and
we can all realize it is also a bit of crock. My AD122 MKII has the
largest dynamic range (127dB unweighted) so it is 21+ bits. My first
generation AD122 had 122dB (20 bits) and I called it a 20 bit
converter. Than all those 100-110dB devices apeared on the market with
"24 bits" on the panel and the sales guys insisted that I call mine a
24 bits. The bits are there, but the last few just bounce around
mindlesly - no realtion to the audio.

I have a cheap circuit for andom number generation. If I add 100 bits
of noie at the thend (least significant bits) will it be a 124 bits
machine? Lets go for it

I could probably design a 24 bit AD, and it will be an unbelivable
energy hog. Remember that we are on a log scale. The differance
between 20 and 24 is not "just 4 more". It is a factor of 16 more.
Sort of like a 6 earthquake vs an 8 earthquake. Big differance.

Do you need 24 bits AD? Probably not, short of some of the headroom
comments. What is the best Mic preamp out there? Say -130dBu? How much
gain is it set for? Say I use 30dB gain, than the noise floor is at
-100dBu and a peak to peak siganl out of the preampp is 24Bu driving
an AD. So we have 100+24=124dB
dynamic range. My AD122 MKII gives you 3dB margin. But say you need
40dB mic pre gain. Now you can use the 114dB dynamic range device... I
am not a recording engineer, but I think that 127dB is already only
for close mics that put out serious signal followed by a great mic
pre...

While I rather have folks specify dynamic range (not bits), the 24 bit
number is pretty good in the sense that it is a multiple of 8 bits
(thus 3 bytes). It is a good number for computers and hardware
(multiple of 8). I just wish 24 bits did not get used by industry
salesman as measure of quality.

Last comment for now: Whatever I said regarding 24 bits and what is
needed is ONLY for AD's and DA's. Let me call them CONVERSION BITS.
There are other type of bits. Let me call them PROCESSING BITS. If 24
conversion bits may be an overkill, 24 processing bits is way too
little. Folks need to realize it. A guy comes with some digital EQ and
tells you that it has 56 bits, or 32 floating point DSP, and he does
have a point to make. Just do not later go and look for a 56 bits
AD... Different issue.

I think it was already explained, but here we go: Say I want to
avarage 50 cent (money) and 51 cent. It is 50.5 cents. I can only deal
with cents so we either call it 50 or 51. We have 1/2 cent inacuracy.
In a bianry wold, I lost a bit of acuracy. So lets agree to have a new
coin - 1/2 cent. Now we avarage 50, 50 and 51. It yields 50.3333333...
Well ae can call it 50 or 50.5.... The general statment is that when
you add more and more computations, you reduce more and more of the
accuracy. So yes, we typicaly need a lot of computational bits.

To all, pardon me if soem of the above is a repeat of what you said, I
tried to fill some gaps, and probably covered things that needed no
help. I am new to audio NG.

BR
Dan Lavry