I have some data that is highly correlated and I wanted to see if I could try and encode it using linear predictive coding (LPC). Here is how I've been understanding the process:


  1. Generate predictive filter coefficients
  2. Filter signal with these coefficients to obtain and error signal $r[n]$.
  3. Check if $r[n]$ is white.
  4. If it is, store the variance of $r[n]$ along with the coefficients and you are done.
  5. If not, increase order until (hopefully) $r[n]$ becomes white.


  1. With the encoding variance, produce a white signal.
  2. Pass that white signal through the inverse filter.
  3. You produce a signal $s[n]$ that has the same autocorrelation sequence and power spectrum as your original.

I've completed all the encoding steps - computed my coefficients, checked if my residual signal was white (autocorrelation is a delta function) and then stored its variance. However, the trouble comes during the decoding process. I took the variance I encoded, produced white noise, and filtered it with the inverse filter. Not only was my signal nowhere near a representation of my old signal, but the two signals had very different autocorrelation sequences as well. I've posted my code and my outputs below. I mainly just want to know if I'm misunderstanding something about LPC or if I'm implementing it wildly incorrectly.

load('amp.mat') %this gets loaded as the variable y with timestamp t

order = 15;
[a,variance] = aryule(y,order);
error = filter(a,1,y);

%Now check to see if error is white

[acs, lags] = xcorr(error,'coeff');
plot(lags,acs), grid;

enter image description here

noise = sqrt(variance)*randn(length(y),1);

%compare noise to error signal

enter image description here

%synthesize signal
yrecon = filter(1,a,noise);
%actual signal, as produced by our residual
yact = filter(1,a,error);

title('Signal vs synthesized signal')

enter image description here

%check autocorrelation sequence of each and plot
[yacs, ylags] = xcorr(y,'coeff');
[yracs, yrlags] = xcorr(yrecon, 'coeff');
plot(yrlags,yracs), grid
title('Autocorrelation sequence of reconstructed signal')
plot(ylags, yacs), grid
title('Autocorrelation sequence of signal'}

enter image description here

  • 1
    $\begingroup$ wow it's exremely efficient to store only the variance of $r[n]$... ;-) consider a signal of length 1 G samples and you reduce it to a few prediction coefficients and a single variance ? something wrong ? $\endgroup$ – Fat32 Nov 25 '18 at 16:02
  • $\begingroup$ I agree that something seems wayyyy off haha. I mean I've tried increasing the number of prediction coefficients, but it neither makes the error 'more white' nor does it seem to make the auto correlation sequences more similar. I apologize, but I'm just really out of my depth here, I've looked all over for examples of implementing LPC for compression but I can't find any worked out examples so I've been trying to do it myself. I just figured that if I can get the residual to be white by removing all correlations in the signal, I've finished the encoding process. I'm guessing that's not true? $\endgroup$ – compscinoob Nov 25 '18 at 19:23
  • $\begingroup$ do you know what LPC is used for and how it does achieve its purpose ? $\endgroup$ – Fat32 Nov 25 '18 at 20:24
  • $\begingroup$ LPC is mianly used for speech because it's designed to fit the model of our vocal system. Our speech is modeled as a filter that is excited by either an impulse train or white noise. I do not know how it achieves its purpose. Although it's used mainly for speech, I don't see why it can't be used for any general signal that's correlated and can be reduced to some residual white noise. Maybe I'm misunderstanding everything, but I would really appreciate it if you could tell me how I could correct my already existing implementation and/or my understanding of LPC. $\endgroup$ – compscinoob Nov 27 '18 at 14:10
  • $\begingroup$ You are ok with the predictive part of the LPC. But it seems you have a problem with the coding part of it; that's where the compression actually takes place. You should consider the quantization stage carefully. I suggest you read Introduction to Data Compression by Khalid Sayood. For a full grasp of LPC and other data compression methods. $\endgroup$ – Fat32 Nov 27 '18 at 15:16

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.