0
$\begingroup$

Why we take different MFCC components in different programs? Suppose, in a program we take No. of mfcc components = 14.What does it signify?

$\endgroup$
  • $\begingroup$ We apply DCT as the last step to obtain MFCCs. DCT is a transformation from filter bank energy coeeficients into cepstrum. It is similar to PCA. We aproximate frequency changes by taking only a few components. The more rapidly a signal of frequency changes, the more DCT coefficients we need. But at some point these changes in frequency are only noise added to signal. Therefore, higher order components of DCT and higher MFCC coefficients will only aproximate noise. This is one of the reasons why we need a limited number of MFCC. How many MFCCs you need depends on application. $\endgroup$ – Celdor Sep 22 '15 at 7:36
1
$\begingroup$

The number of coefficients is just one of the many parameters to tweak when designing a classifier. For speech analysis it turned out that usually 8-14 coefficients are sufficient, and increasing the number of coefficients beyond 14 does not improve the performance for most applications. The exact number you choose will depend on your task at hand. Most systems I've worked with used 12 coefficients, and that was for speech recognition. If you're not sure, you can simply run some experiments and change the number of MFCCs in a reasonable range to see the effect it has for your application. But I have no doubt that you'll end up in the range that I've mentioned above.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.