[net.audio] CD digital filters-Eidson errs.

jj@alice.UUCP (08/25/85)

>From allegra!ulysses!burl!clyde!bonnie!akgua!whuxlm!harpo!decvax!linus!philabs!prls!amdimage!steve Wed Dec 31 19:00:00 1969

> 	It's almost amusing listening to two guys from
>the same company throw things at eahc other, but...

Wrongo.  CI-2 says otherwise, budget brain.

>	First off, in every design case I have encountered, 
	... lots of semi-true stuff
>	The other thing that has been ignored completely
>is that most output interpolation filters I have experienced
>are linear phase FIR filters. The response of a linear 
>phase FIR filter has no ringing. The most common filter

Oh wow!  What a jem.  "Linear Phase FIR filters have no ringing"  Well,
sir, may I recommend any simple, first year text on
either signal processing or communications.  Either will show
you the complete untruth of your statement.  The ringing must be
symmetric, indeed, but ringing there is, indeed 

>I have used for interpolation is the simple: 
>	y(n) = 0.25*x(n) + 0.5*x(n-1) + 0.25*x(n-2). 	
What is its frequency response?  How well does it anti-alias?
<I know, you tell ME!  You've picked the simplest filter of the
bunch.>

Well, the model for the ideal interpolation filter is 
the sync function (sin x)/x, which has lots of zeros, more and more
as you use longer versions, and which clearly rings all the way
to infinity.  You don't have to use it all, though, so the ringing
is time-limited, at least.  (I mean FIR filter, too.  It really doesn't
matter, though, since to get an arbitrary amount of frequency-amplitude
resolution, the significant part of the impulse response will be the
same length for FIR, IIR, or even analog.  A simple fact of nature, math,
and life in general.)

>Zeros are alternately inserted into the sequence to get a 
>factor of two increase in the output data rate, and the output 
>is multiplied by 2 to produce a properly scaled output.

You're half right.  Figure out which half yourself.

>	It should also be noted that digital filters
>exhibit sensitivity to coefficient truncation, just as
>analog filters are sensitive to component selection. Since

FIR filters?  Coefficient truncation?  Well, yeah, sort of,
but in a totally different way that's much easier to deal with,
and which can be modeled when you design the filter, permitting
you to design the best filter INCLUDING truncation...

>multipliers are still relatively expensive for a consumer
>product such as a CD player, I assume that most manufacturers
>use a simple shift-and-add scheme. The example I gave above
Nope.
>can be implemented on most off-the-shelf CPUs. Many IIR 
At a sampling rate of 1kHz, maybe.
>filters (Butterworth, Chebyschev, and elliptic) have
>poles near the unit circle, which can be translated into 1-2^n,
>but they are not as simple as the one-bit coefficients in
>the FIR interpolation filter. I suggest that an IIR digital
sigh.  wrong. figure it out yourself.
>filter may appear to have the same response as its analog 
>counterpart in an infinite precision computer simulation, but
>in a real implementation, an digital filter will exhibit sensitivity
>to coefficient truncation, while an analog filter will be
>sensitive to the component values selected. 
True completely for IIR designs.
>
>	Another reason for choosing a digital interpolation
>filter, besides time and temperature stability, is the exact
>replication of a given filter from CD player to CD player.
Amen!
>This coupled with the fact that the analog anti-alias 
>filter can have much more compnent "slop", weighs in favor of the
>oversampled digital approach rather than an analog approach
>on mass produced players (emphasis on mass).
agree here, except it's true for both MASS and small runs.  It's 
true, period!  Of course, there are other problems, but they're
not what you're discussing.  The analog filter sensitivity and
non-deterministic, continuous "coefficient values" will always
get you.  No matter how hard you work.
>Sorry about the long-winded-ness, but I had to absolve myself
>of this information. Any intellectual discussions are gladly
>invited, any impassioned mumbo-jumbo to /dev/null.
Please go read a good communications text, and then come back and
make your explainations again.  Rabiner and Gold, Rabiner and Shaeffer,
and many other books will quickly disabuse you of some of your
rather random comments.


This is just another proof of why audio is such a rotten field to
work in.  I agree completely with Dick Pierce(sp) in that.
Arrogance, indeed, to suggest that "Linear phase filters have
no ringing".  
-- 
SUPPORT SECULAR TEDDY-BEAR-ISM.
"You, who are on the road, must have a code that you can live by."

(ihnp4/allegra)!alice!jj

steve@amdimage.UUCP (Steve eidson) (08/28/85)

In article <4213@alice.UUCP> jj@alice.UUCP writes:
>>are linear phase FIR filters. The response of a linear 
>>phase FIR filter has no ringing. The most common filter
>Oh wow!  What a jem.  "Linear Phase FIR filters have no ringing"  Well,
>sir, may I recommend any simple, first year text on
	First, I'd like to apologize for making this statement
sound like a theorem. Of course FIR filters exhibit ringing, but
the interpolation filter I was referring to in the article does not.

>>I have used for interpolation is the simple: 
>>	y(n) = 0.25*x(n) + 0.5*x(n-1) + 0.25*x(n-2). 	
>What is its frequency response?  How well does it anti-alias?
><I know, you tell ME!  You've picked the simplest filter of the
>bunch.>
	If you knew what the frequency response was, why did 
you make this comment. You and I both know it is a low-pass
filter and the frequency response is far too complex to
broadcast over the net.

>Well, the model for the ideal interpolation filter is 
>the sync function (sin x)/x, which has lots of zeros, more and more
	Somebody else should check their elementary communications
text, it's "sinc" function.

>>Zeros are alternately inserted into the sequence to get a 
>>factor of two increase in the output data rate, and the output 
>>is multiplied by 2 to produce a properly scaled output.
>You're half right.  Figure out which half yourself.
	You better figure it out, it is all correct.

>>poles near the unit circle, which can be translated into 1-2^n,
>>but they are not as simple as the one-bit coefficients in
>>the FIR interpolation filter. I suggest that an IIR digital
>sigh.  wrong. figure it out yourself.
	you're wrong. you figure it out.

>>filter may appear to have the same response as its analog 
>>counterpart in an infinite precision computer simulation, but
>>in a real implementation, an digital filter will exhibit sensitivity
>>to coefficient truncation, while an analog filter will be
>>sensitive to the component values selected. 
>True completely for IIR designs.
	True completely for all digital filters.

>Please go read a good communications text, and then come back and
>make your explainations again.  Rabiner and Gold, Rabiner and Shaeffer,
>and many other books will quickly disabuse you of some of your
>rather random comments.
>This is just another proof of why audio is such a rotten field to
>work in.  I agree completely with Dick Pierce(sp) in that.
>Arrogance, indeed, to suggest that "Linear phase filters have
>no ringing".  

	Sir, I have no quarrel with you, but the overall general
hostile tone of your reply to my article was totally uncalled for.
I posted the article hoping to earnestly enlighten some of those
who are not familiar with DSP. How can honestly call me arrogant
after your reply. I admit my mistake, so please tone down your
conversation.
 ----------
 "...but you've got no arms and no legs,
 what are you going to do, bleed all over me ..."

 Steve Eidson (408) 749-2303
 UUCP: {ucbvax,decwrl,ihnp4,allegra}!amdcad!amdimage!steve
 ARPA: amdcad!amdimage!steve@decwrl.ARPA

steve@amdimage.UUCP (Steve eidson) (08/28/85)

	One other point I forgot to mention, it is possible 
to implement interpolation filters in shift-and-add architectures.
If the coefficients are kept simple, very high sample rate
filters can be implemented. The FIR interpolation filter
argued about previous has been implemented (by me at AMD)
in a shift-and-add architecture at a 128kHz sample rate.
We have also used more complex filters in the same architecture
at sub-multiples of this sample rate. 
	Unfortunately, many people have developed the notion
that they MUST have a multiplier to do digital signal processing,
when there are other alternatives for simple systems. Perhaps
none of the filters I have just mentioned would be adequate
for a CD player, I really haven't gone through the process of
designing an interpolation filter for CD's.
	In any case, I would be glad to discuss the details of the
interpolation filters I have talked about here with anyone.
Please send mail, I don't want to clog the net.

Thanks for your patience.
 ----------
 "...but you've got no arms and no legs,
 what are you going to do, bleed all over me ..."

 Steve Eidson (408) 749-2303
 UUCP: {ucbvax,decwrl,ihnp4,allegra}!amdcad!amdimage!steve
 ARPA: amdcad!amdimage!steve@decwrl.ARPA