I want to normalize my data prior to a deep neural network model (DNN). It was recommended to me to use real mean square (RMS) normalization for audio, though I am not sure this is the best for a DNN. I am using MFCC features to represent the data.

Is RMS a good method? If so, how do I choose a good RMS value to normalize with respect to that value? Is there a rule of thumb?

  • $\begingroup$ Are you planning on using 0'th MFCC coefficient (energy)? $\endgroup$
    – jojek
    Nov 15 '20 at 12:29
  • $\begingroup$ If I understand you correctly, I am using the lowest MFCC line (was that what you were referring to?) $\endgroup$
    – havakok
    Nov 15 '20 at 13:13
  • $\begingroup$ You will calculate certain number of coefficients for each frame of audio (be it 13, 16, etc). Are you gonna use them all? The 0'th coefficient (depending on implementation) will be the energy of all log-mel energies. $\endgroup$
    – jojek
    Nov 15 '20 at 14:29
  • $\begingroup$ I am using MFCC with 40 coefficients and I am using all of them. $\endgroup$
    – havakok
    Nov 16 '20 at 7:41

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.