From: John Gilmore <gnu@toad.com>
Date: Wed, 5 Jun 1996 12:37:04 +0800
To: cypherpunks@toad.com
Subject: MELP: 2400 baud speech coding
Message-ID: <199606042031.NAA25407@toad.com>
MIME-Version: 1.0
Content-Type: text/plain


Mixed-Excitation Linear Predictive encoding gives better speech
quality than CELP at half the data rate.  Encoding and decoding
together burn up more than 100% of a TMS320C3x digital signal
processor at 33MHz -- 64% to encode and 53% to decode.  I don't know
how it does on a Pentium or an Alpha.  If you have the MIPS at both
ends, this enables very robust encrypted speech across modem links to
the Internet.  John Walker's free SpeakFreely software
(http://www.fourmilab.ch/speakfree/windows/speak_freely.html; or
follow the link from there to the Unix version) is already doing
packet replication for high reliability, using the earlier LPC-10
algorithm.  (It doesn't implement MELP, though those on fast CPUs
could add it.)

I noticed an ad in EE Times that said, "MELP: The new Federal Standard
for 2400 bps Speech Coding", so I did a web search for it.  It
reportedly comes from Georgia Tech research.  Atlanta Signal
Processors has the exclusive license.  See
http://www.ti.com/sc/docs/dsps/softcoop/voc-13-1.htm.

	John Gilmore