Newsgroups: comp.speech
Path: lyra.csx.cam.ac.uk!doc.ic.ac.uk!agate!howland.reston.ans.net!europa.eng.gtefsd.com!MathWorks.Com!news2.near.net!noc.near.net!analog.com!analog.com!nwd2sun2.analog.com!az
From: az@saturn.analog.com (Alex Zatsman)
Subject: Re: LPC synthesis (What? How?)
In-Reply-To: Pete@adder.demon.co.uk's message of Sat, 20 Aug 1994 16:14:55 +0000
Message-ID: <AZ.94Aug23104945@saturn.analog.com>
Sender: usenet@analog.com
Organization: Analog Devices Inc, Norwood MA, USA
References: <330245751wnr@adder.demon.co.uk>
Date: Tue, 23 Aug 1994 15:49:45 GMT
Lines: 29

In article <330245751wnr@adder.demon.co.uk> Pete@adder.demon.co.uk (Pete Rowley) writes:

>    Could someone please either explain in reasonably simple terms what LPC 
>    is and how it works, or direct me to literature either on the net or 
>    not that could help me.

Let me try. Say you have a signal which  is a mix (linear combination)
of 2 sine waves (harmonics). Then you can generate the signal by using
all-pole (IIR) filter, or recursive difference equation:

	y_n = a_1*y_{n-1} + a_2*y_{n-2} + a_3 * y_{n-3} + a_4 * y_{n-4}

These 4 numbers together with 4 initial conditions (excitation vector)
can describe the signal.  If  the signal has N harmonics,  you need 2N
coefficients and 2N initial conditions.

This is  very convenient for speech  processing because speech signals
can usually be approximated by signals with  no more than 5 harmonics.
By computing   just 10  coefficients    you   get a   relatively  good
approximation,   and this  can  be    used,   for example, in   speech
compression. It seems that  the main difference between various speech
compression schemes is  in the way they  describe the excitation, i.e.
CELP   for Codebook Excited  Linear   Prediction, VSELP for Vector Sum
Excited LP, etc.

Hope this did not confuse too much.
--
Alex Zatsman,     Analog Devices, Inc., (617) 461-3729
alex.zatsman@analog.com
