From: gnu (John Gilmore)
Date: Fri, 5 Feb 93 11:17:05 PST
To: pmetzger@shearson.com (Perry E. Metzger)
Subject: Re: `Sunday Times' article on GSM changes
In-Reply-To: <9302051715.AA22381@maggie.shearson.com>
Message-ID: <9302051916.AA23983@toad.com>
MIME-Version: 1.0
Content-Type: text/plain


> > ... voice data requires about 64Khz...

Actually I think you meant 64 kbits/sec.

> Just as a point of information, you can compress voice into 4kbit/sec
> without any real loss of quality.

I wouldn't say "without any real loss of quality".  I have an AT&T secure
phone on loan for evaluation (attendees at the last Mt. View cypherpunks
meeting have played with it).  It compresses voice to 4800 b/s then encrypts
it and uses a stripped down V.32 modem to send it.

The voice is intelligible on the other end, but there is definitely a
loss of quality.  It sounds like you are talking through the bubbler
in an aquarium (underwater).  But it is possible to recognize
someone's voice on the other end, at least when you already know who
it is.  "No real loss of quality" goes way too far.

By the way, I have public domain (federally written) code that
compresses voice down to this size.  It's called CELP, Code Excited
Linear Prediction.  The only catch is the code they wrote runs much
slower than realtime (on workstations).  My guess is that there are
significant speedups that we could make by hacking on it and running
gprof.  Van Jacobson has done some work on this, but his policy seems
to be to sit on anything good for two to three years before releasing
it.  Anyone interested in beating him to the net with something that
would compress voice (or voicemail) in realtime on a 486 or a SPARC?

	John