Re: AUDITORY Digest - 26 Sep 2010 to 27 Sep 2010 - Special issue (#2010-228)

[karan behar <karanbehar@xxxxxxxxx>]

Dear Sir

Greetings. I apologize for asking mundane and trite questions in my previous posts. Thank you very much for your suggestions. 

I wanted to ask a question on Mel Frequency Cepstral
 Coefficients for obtaining feature vectors. In many papers I have read 
about people obtaining 13 coefficients and in some about obtaining 12 MFCC using FFT and then using them to treat the network. The example of such paper is Logan B. "Mel frequency cepstral coefficients for music modeling", International Symposium of Music Information and for 12 coefficients Fukada T.  "An adaptive algorithm for mel-cepstral analysis of speech", ICASSP, 1992. 


1. Why do we need 13 or 12 coefficients. What are the differences between them. 

2. I am having hard time to understand the what each of the thirteen coefficients physically and mathematically  represent for MFCC 13 and 12 coefficients. For example if first coefficient represents frequency and second represents timber etc. 

I have been using 13 MFCC to train my network but I do not know everything that works behind using 13 or 12 coefficients.

Through and fundamental papers addressing the mathematics and significance of these questions will be very helpful.


Please advise.

Regards

Karan