I have two sets of MFCC coeficients which I get with fftExtract utility from two wav audio files.

How can I compare this sets (matrix 271x13 and 551x13) if they are similar ?

Probably I should use EMD distance ... But Im sligthly lost in all this math and I dont know how to interpret its results .

How can I make from this distance matrix if both audio signals are similar or not - (Yes/Not)

I used following code for EMD :

function D = distEmd( X, Y )

Xcdf = cumsum(X,2);
Ycdf = cumsum(Y,2);

m = size(X,1);  n = size(Y,1);
mOnes = ones(1,m); D = zeros(m,n);
for i=1:n
  ycdf = Ycdf(i,:);
  ycdfRep = ycdf( mOnes, : );
  D(:,i) = sum(abs(Xcdf - ycdfRep),2);



  • 2
    $\begingroup$ any specific reason you are using earth mover's distance? Wiki says it is used to find distance between probability distributions. I don't think MFCCs are prob. distributions. Perhaps you could use Dynamic Time Warping method to find distances. $\endgroup$ – user13107 Aug 22 '13 at 16:38
  • $\begingroup$ Well I have read some articles where people used it with MFCCs so I wanted to try it . I think EMD is faster than DTW. link $\endgroup$ – Mark Kica Aug 23 '13 at 17:40
  • $\begingroup$ I don't think EMD is a good idea for audio signals, faster it may be, but its idea is to compare distributions/histograms. Anyway, if you want to use it then create a 50 bin histograms from your data. Now both matrices will be of size 50x13. $\endgroup$ – user13107 Aug 24 '13 at 17:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.