Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.1 6/24/83; site drune.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!mhuxr!mhuxt!houxm!mtuxo!drutx!drune!mohler
From: mohler@drune.UUCP (MohlerDS)
Newsgroups: net.audio
Subject: Re: David Mohler is completely correct
Message-ID: <24@drune.UUCP>
Date: Thu, 22-Aug-85 22:52:22 EDT
Article-I.D.: drune.24
Posted: Thu Aug 22 22:52:22 1985
Date-Received: Sun, 25-Aug-85 04:04:28 EDT
References: <19@drune.UUCP> <4162@alice.UUCP>, <10068@ucbvax.ARPA>
Organization: AT&T Information Systems Laboratories, Denver
Lines: 44

Steve,
	Thanks for the pleasant and accurate reply, it is refreshing that
not everyone uses flames to clarify or dispute an article.

	I must appologize for the haste in which my initial points were
made, they suffered seriously as a result. I agree that a 16 bit quantizer
improves S/N by 12db and that 4x oversampling only yields a 6db improvement
over a simple 14 bit quantizer. My point was that you are close to 16 bit
performance, but it is far more accurate, as you say, to state clearly that
you have 15 bit performance with respect to noise! 

If however, you use noise
shaping ( per the philips scheme ) where the 28 bit output of the transversal
filter is rounded off to 14 bit data, then the 14 LSB's are delayed by
one sampling period, reversed in sign, and summed with the next sample
the result is a 7db improvement in the average quantization error and noise
for low frequency signals (below 22.05KHZ). The philips scheme also uses
a 3rd order low pass bessel filter that is 
3db down at 30khz. All of this means that
from DC to 22.05 KHZ you have a 1db better S/N with the philips 14 bit
system than a non-oversampled 16 bit system. 
This I contend is virtually the same as
16 bit performance with respect to signal, noise and linear amplitude, and
was my original point!

I also agree that the noise performance is not based on the filter type
(analog or digital), but instead the size ( number of bits ) of the
quantizer, the oversampling ( if any ), and the noise shaping ( again, if
any ). My other point was since this lets you use a filter that is easier
to implement in the consumer electronics price range and since it allows you
to use an ADC that is cheaper (in some cases) than a 16 bit ADC that it
seems to me, the better approach. The cost of a 24th order filter (The 
theoretical 96th order transversal filter, 
altered by the fact that since data only
arrives at the sampling rate which means that 3 out of 4 of the numbers
multiplied in the filter would be zero, therefore the filter can be simplified
to 24 elements.) done digitally, both in terms of power and dollars
will be less expensive than an analog filter.

Thanks again for the comments!

			David S. Mohler
			AT&T - ISL @ Denver
			drune!mohler or druxu!mohler