Newsgroups: sci.electronics
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!gatech!howland.reston.ans.net!torn!nott!bnrgate!bnr.co.uk!uknet!cam-eng!cmh
From: cmh@eng.cam.ac.uk (C.M. Hicks)
Subject: Re: 8x oversampling CD player
Sender: cmh@eng.cam.ac.uk (C.M. Hicks)
Message-ID: <1993Apr22.111443.9703@eng.cam.ac.uk>
Date: Thu, 22 Apr 1993 11:14:43 GMT
References: <hcbC5un9L.DD0@netcom.com> <1993Apr21.160303.10838@doug.cae.wisc.edu> <mcmahanC5v942.MKJ@netcom.com>
Nntp-Posting-Host: club.eng.cam.ac.uk
Organization: cam.eng
Keywords: oversampling, CD, digitized audio
Lines: 82

mcmahan@netcom.com (Dave Mc Mahan) writes:

> In a previous article, kolstad@cae.wisc.edu (Joel Kolstad) writes:
>>In article <hcbC5un9L.DD0@netcom.com> hcb@netcom.com (H. C. Bowman) writes:
>>>
>>>I just bought a new portable CD player for the office, and I notice that
>>>it proudly proclaims "8 TIMES OVERSAMPLING" on the box.  Now while I think
>>>I understand what oversampling is (the rate of discrete "samples"
>>>exceeds the highest frequency component of interest by some factor),
>>>I don't understand this "8 TIMES" business...  It seems to me that when
>>>I bought my first CD player (was it REALLY 10 years ago?!), the specs
>>>said "4 TIMES" ...  Could someone please tell me whether I'm getting
>>>senile?  If I'm not, then what good does it do for the player to take
>>>samples at a higher rate?  

>>The data is only ever read once (barring mistracks and such, of course),
>>and eventually gets turned into 44.1 KHz, 16 bit, two channel data.
>>Oversampling takes two discrete data points, and interpolates n-1 points
>>between them for n times oversampling.  When I asked, people said that the
>>interpolation was not simply linear interpolation, but significantly more
>>complicated.

>You are quite correct in your understanding.  The filtering is not
>interpolation, as that would distort the frequency content of the signal

They do interpolate - it's not linear interpolation though. You are correct
that the frequency content is not altered (more specifically, the baseband
spectrum is preserved, and so is every Nth image spectrum. The other
(N-1) in N image spectra are removed, where N is the oversampling rate)

>you are listening to.  Generally, these players run the samples thru an
>all-pass filter network.  I have done this for ECG waveforms from a person's

Ideally the filter has a gain of 1 from 0Hz to 22050Hz, and a gain of 0
from 22050 to the new Nyquist frequency. In practice a finite transition
band is required, and there is also a certain amount of pass-band ripple
and stop-band leakage. With a high order (eg 200 taps) digital filter,
a very good approximation can be easily achieved.

 <<< BIT DELETED ABOUT OBSERVATIONS OF INTERPOLATION FILTER AT WORK >>>

>In case you care, the filtering method uses an FIR (finite impulse response)
>filter.  I'd guess that CD makers use the same kind of method.  Anybody out
>there know the real answer?  I'd say that they use a tapped delay line with
>resistor/op-amp weighting to accomplish the filtering.  This strikes me as
>the most cost effective method for volume production runs.

No, they actually use a digital FIR just like yours, but built in hardware,
on custom VLSI chips for (mainly) economic reasons.

>>Anyway, then, the purpose of oversampling is to move the "effective"
>>sampling rate up to n times 44.1 KHz, in order to use higher frequency
>>antialiasing filters.  For the same quality filter, higher oversampling
>>lets you build cheaper filters, whereas for the same price filter, higher
>>oversamplings lets you build better filters.  So, assuming the quality of
>>all other components in a CD player remained the same, oversampling should
>>allow a manufacturer to produce _slightly_ better sound due to anti-alias
>>filtering.

>Actually, I think the only reason they do this is so that they can say that
>they have a marketting gimic.  I would guess that it is acutally cheaper to
>filter an oversampled signal than not.  You can use sloppier components and
>give the filter a roll-off that isn't so sharp.   It's too bad that they
>charge more for something that (I think) is actually less costly to build.

I guess that the answer is somewhere between the two.

As an interesting (?) aside, some of you out there may be aware of a scheme
used by Pioneer and Wadia, which is called (by Pioneer, at least) Legato
Link. I haven't heard one of these players, but by reading advertising
blurb, and a couple of pseudo-technical articles I deduced that all they
do is build a lousy digital oversampling filter, and let a load of the
first image (ie aliassed) spectrum through. Talking to Bob Stuart (of
Boothroyd Stuart, aka Meridian) confirmed my suspicion. He said that it
sounded awful, but then he would, wouldn't he...

Christopher
--
 ==============================================================================
  Christopher Hicks    |      Paradise is a Linear Gaussian World
  cmh@uk.ac.cam.eng    |    (also reported to taste hot and sweaty)
 ==============================================================================
