From: "Dave Emery" <die@pig.die.com>
Date: Thu, 1 Dec 94 15:03:36 PST
To: eric@remailer.net (Eric Hughes)
Subject: Re: recent voice over data
In-Reply-To: <199412012150.NAA13820@largo.remailer.net>
Message-ID: <9412012302.AA02541@pig.die.com>
MIME-Version: 1.0
Content-Type: text/plain


> 
> 
>    the great voice-over-data protocols and products introduced by
>    Intel, Rockwell, ZyXEL and others at Comdex which will make
>    Voice-PGP so much easier
> 
> As I understand these voice-over-data products, the voice goes over
> analog, added to the modem signal.  The modem signal is interpreted,
> and then reconstructed and subtracted from the incoming signal,
> leaving voice.  Very clever, but insufficient for secure phones.
> 

	I can't quite see how this would work unless the voice was run
at a very low level relative to the data.  In order to subtract the
modulated version data coming from the other end you have to know
exactly what it is, and in order to do that you need sufficient signal
to noise of data over everthing else to reliably demodulate it or you
need some means of reliably predicting it.
	
	  Now I recognize that some of the time there is little or no
entropy in the information in one direction (it is completely
predictable - such as flags during LAPM idle intervals) and it is
possible that one could contruct a syllabic gizmo that would turn off
the entropy in the data when talkspurts happened by doing flow control
and stopping information transmission during periods that speech was
loud.  But all of this seems a bit much, and certainly would be subject
to lots of kinds of degradation depending on the speech content and
any nolinearity in the channel.

	In general the modulations used in modems require at least 12-15 db
of SNR for decent BERs - this would imply that if the line was not
timeshared with voice in talkspurts that the peak level of the voice
would have to be about 20 db below the modem tones.  I guess that this
would still result in intelligible speech even though it would effectively
be transmitted by only a couple of bits per 8 khz sample.

	And yes I guess that by using forward error correction on the
data at a variable coding rate one could allow the voice to creep
up on the data a bit more by using a heavier duty error correction
during talk spurts.

	I had assumed that these products digitized voice and multiplexed
it with the data stream however.   Certainly there are technologies to
do this including variable rate vocoding (ala Qualcomm) that would
allow almost the full bandwidth of the line to be used for data.


	[ And yes I'll bring up premail tommorow so I can sign things if
I can get around to it without interrupting serious work work - but I'm left
wondering who in the hell would want to forge posts from me ? ]

						Dave Emery N1PRE