NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/rec/audio/10882

Internet Message Format | 1992-07-27 | 3.9 KB

Path: sparky!uunet!walter!att!att!allegra!alice!jj
From: [e-mail] (jj, curmudgeon for trvth in audio)
Newsgroups: rec.audio
Subject: Re: Compressed audio better than noncompressed?
Message-ID: <[e-mail]>
Date: 27 Jul 92 14:20:35 GMT
Article-I.D.: alice.23326
References: <[e-mail]> <[e-mail] u.au> <[e-mail]> <[e-mail]> <[e-mail]>
Reply-To: [e-mail] (jj, curmudgeon for trvth in audio)
Organization: NJ State Home for Bewildered Terminals
Lines: 69

In article <[e-mail]> [e-mail] (Stuart Woolford) writes:
>[e-mail] (Peter Monta) writes:
>
>>In article <[e-mail]> [e-mail] (Curt Welch) writes:
>>> My guess is that you can probably do something like 8:1 loss free
>>> compression on CD's but designing and building the hardware to do this
>>> in real time is currently too expensive.
If by "loss-free" you mean "perceptually loss-free", some algorithms
have been controversially verified by the ISO-MPEG-Audio committee
to be "transparent" at 2.66 bits/sample, when compression 16 bit samples.
Nobody I'm aware of (and I do study the field a wee bit) has reported
transparency <including myself, to the present> below that
rate, although some algorithms (PAC, ISO-Layer III, for instance,
there may be others) may provide such performance.

>>If you really mean "loss free", i.e., the original 16-bit data is
>>reproduced, then this is impossible.  Any audio recording on CD has
>>an entropy larger than 2 bits/sample.

>why?? ( no, really, I would like to hear you're reason... )
Well, because many people have measured it.  If you'd like some
confirmation, look in Chapter 4 of the book "Advances in Speech
Signal Processing" by Furui and Sondhi.  Chapter 4, written
by James Johnston of AT&T Bell Labs, and Karlheinz Brandenburg
of the University of Erlangen, FRG, is titled "Wideband Coding-
Perceptual Considerations for Speech and Music".

On Page 110 and 111, they list some LPC gain statistics and Spectral
Flatness Measure (SFM) statistics that show a largest SFM
gain of 45.6 dB at a 512 sample Hann Window.  This, if you
consider its entropy considerations, means that about half of the
bits in the signal are actually information-carrying, i.e. you can
encode this signal in about 8 bits, give or take, and wind up
with an exact reconstruction.

Other work suggests that for most audio, 5-7 bits/sample is the known
limit for exact reconstruction, using the most modern techniques.

If one measures the single-sample entropy in a common audio signal,
perhaps the first 20 seconds of "Tom's Diner" from Solitude
Standing (Susan Vega), one might find that the single-sample entropy
is about 8.3 bits/sample.  This, of course, doesn't remove
intersample redundancy.

Good luck using LZW.
>
>>> ... As I understand it, both the DCC and MD compression algorithms
>>> eliminate some sounds based on the human ears ability to here them
>>> because of masking effects.
>>> ...
>>> So, I'm just wondering when people will start claiming that
>>> compressed audio sounds better than non-compressed becauses
>>> it has been purified?
>
>>Enhancement isn't the goal of these algorithms; they attempt to
>>minimize the loudness of quantization noise, so to the extent that
>>the compressed audio is indistinguishable from the non-compressed,
>>these algorithms are good.  Enhancement/restoration is an interesting
>>topic, but it's separate from source coding.
>
>>Peter Monta   [e-mail]
>>MIT Advanced Television Research Program


--
 Bush    *Copyright alice!jj 1992, all rights reserved, except transmission
   vs       *by USENET and like free facilities granted. Said permission is
 Clinton  --    *granted only for complete copies that include this notice.
 Just Say !NO!       *Use on pay-for-read services specifically disallowed.