steve@amdimage.UUCP (Steve eidson) (08/29/85)
I'm not trying to belabor a point, but I thought I'd
post a follow-up now that I have some information on real
CD interplation filters. (At least before someone tries to
beat me about the head and shoulders again. :-)) The source
of this info is the article "Communications Aspects of the
Compact Disc Digital Audio System" from the Feb. 1985 issue
of the IEEE Communications Magazine. The author, Dr. J.B.H.
Peek is with Philips Research Labs.
According to Dr. Peek, their interpolation filter is a
96-tap FIR filter. The input to the filter is 16-bit, and 3
zero samples are inserted between each input sample to increase the
sample rate by a factor of 4. The filter has 12-bit coefficients,
but the article has conflicting statements as to whether 24- or
28-bit arithmetic is used in the filter. A guess would be 24-bit
because 12 by 12 multipliers are available (although a 16 by 16
could just as easily be used). For this length and complexity
filter it is necessary to have a multiplier. Because of the inserted
zero samples, only 24 multiplies are required per filter update.
I assume there is also a 24-bit adder to round out the filter
hardware. The filter passband is 0-22 kHz and the the stopband
is 24-88.2 kHz. The article does not state what the passband
ripple is, but the stopband attenuation is 50 dB.
The output of the filter is rounded to 14 bits, which
would have 6dB disadvantage compared to its 16-bit un-interpolated
counterpart (interpolatation by 4 provides an effective 15-bit
resolution). Then, there's a trick that Philips uses. They
apply a noise shaping filter (no details provided) to the
interpolator output. The noise shaping filter reduces the in-band
(0-20kHz) noise at the expense of increased out-of-band noise.
The article claims that this reduces the audible noise by another
7 dB thus making the interpolated system of comparible performance
to the 16-bit system.
As I stated in a previous posting, the interpolation
process makes the analog anti-alias filter much simpler than
the 44.1 kHz sample rate / 16-bit system. This, along with
the replicability of digital filters from unit to unit argues
in favor of the interpolative approach. The point I was trying to
make with my original posting (I did a rather poor job) was
that the nature of FIR filter implementation allows a reduction
in the computational power required compared to its IIR
counterpart. (Yes, I know the order of the IIR filter would
be much lower than the FIR filter order.) The FIR filter
allows an EFFECTIVE computational rate equal to the input
sample rate because only 1 out every 4 multiplies needs to be
performed. An equivalent IIR filter must be updated at the
output sample rate due to the feedback involved.
Thanks again 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.ARPAmalcolm@spar.UUCP (Malcolm Slaney) (09/01/85)
In article <519@amdimage.UUCP> steve@amdimage.UUCP (Steve eidson) writes: >Then, there's a trick that Philips uses. They >apply a noise shaping filter (no details provided) to the >interpolator output. The noise shaping filter reduces the in-band >(0-20kHz) noise at the expense of increased out-of-band noise. >The article claims that this reduces the audible noise by another >7 dB thus making the interpolated system of comparible performance >to the 16-bit system. I puzzled over this for a while after I read it and decided that they must be doing dithering. Revox published a white paper on their CD player and described the technique. I don't remember whether the SNR improvement is correct. Cheers. Malcolm
sjc@angband.UUCP (Steve Correll) (09/03/85)
> In article <519@amdimage.UUCP> steve@amdimage.UUCP (Steve eidson) writes: > >Then, there's a trick that Philips uses. They > >apply a noise shaping filter (no details provided) to the > >interpolator output... > I puzzled over this for a while after I read it and decided that they > must be doing dithering. Philips may also use dithering, but their noise-shaping is something else. Because Philips rounds each 16-bit sample to 14 high-order bits before presenting it to the D/A, circuitry prior to the D/A accumulates the error from each sample (that is, "16-bit sample" minus "16-bit sample rounded to 14 bits") and, when the error becomes big enough, subtracts it from the next 14-bit number being sent to the D/A converter. Thus, over a time scale of many samples, the integrated output of the D/A is the same as if the D/A accepted 16 bits. Dithering is a well-known technique which adds noise or random values to a signal to reduce digitization errors at the cost of S/N ratio. One cheap way to anti-alias an image for display on a raster monitor, for example, is to add noise to the edges of things. Anyone interested in the technology of the CD will want to read not only the paper by Peek in the February 1985 "IEEE Communications Magazine", but also Volume 40 No. 6 of the "Philips Technical Review" (1982), a collection of several papers on the CD. They are a giant step above what usually passes for technical information among audio manufacturers. -- --Steve Correll sjc@s1-b.ARPA, ...!decvax!decwrl!mordor!sjc, or ...!ucbvax!dual!mordor!sjc