Calculating MFCC
in [DSP]
|
From: Iris.Adam on 24 Jan 2010 08:27 HI everybody... I'm brand new to this community and also new to DSP, and this is my first time asking questions in any community so forgive me for my long questions, I have some questions I appreciate if you could help me. I calculated MFCC but I think it is incorrect i'll explain the way i did it so if you can tell me if it is correct or not,and if not how to fix it. Actually I don't understand it well, although i read many articles about it. In the steps after pre-emphasizing the data, for example windoing, fft etc. i take it for each block like this - pre-emphasize all the signal - for each overlapped frame do take hamming window, and put result in array (h) take FFT for the data in (h) find magnitude of the fft data apply triangular window to the magnitude data find summation of the previous step find log of the summation apply DCT to the log data (this will be the cepstrum) next overlapped frame - calculate Nth coefficient by summing Nth value of each frame - then find Delta and DeltaDelta by these two equations DeltaC(k)= C(k+2) - C(k-2) DeltaDelta(k) = C(k+1) - C(k-1) Is this correct or just I have to apply the hamming to the whole data instead of each frame and then calculate Fourier? Please help....... I have some other questions: 1- How to decide number of triangular windows in the mel scale step, it depends on what? 2- the equation of mel scale mel= 2595 + log(f/ 7000 + 1) where will be used, in the articles I read i didn't understand when to use it, is it used before triangular windowing? 3- In calculating triangular windows, when says if m[f-1] > k and k < m[f] what is k? is it the frequency values(ones we obtained from fft) or it is just the iteration control variable? because when i multiply this to the magnitude value and find the summation I get negative values, which means its wrong 4- is there any way to ensure that the results are correct 5- this one is about NN. how to determine number of inputs to the NN ( I use back-propagation NN)
From: Jerry Avins on 24 Jan 2010 09:25 Iris.Adam wrote: > HI everybody... > I'm brand new to this community and also new to DSP, and this is my first > time asking questions in any community so forgive me for my long questions, > I have some questions I appreciate if you could help me. > > I calculated MFCC ... You left something out. What is MFCC? Read http://users.erols.com/jyavins/procfaq.htm Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
From: Richard Dobson on 24 Jan 2010 11:54 On 24/01/2010 14:25, Jerry Avins wrote: > Iris.Adam wrote: >> HI everybody... >> I'm brand new to this community and also new to DSP, and this is my first >> time asking questions in any community so forgive me for my long >> questions, >> I have some questions I appreciate if you could help me. >> I calculated MFCC ... > > You left something out. What is MFCC? Read > http://users.erols.com/jyavins/procfaq.htm > > Jerry Mel-Frequency Cepstral Coefficients. Form of filtering. Used in speech recognition especially, but also in sound description and classification generally, e.g. MPEG-7 'n stuff. It has been discussed before here. Google is... etc. Seems to me the OP's question ticks all the relevant boxes in that faq; but the faq doesn't address the issue of what dsp topics/acronyms must first be explained to the group (again). That said, there may well be more people on the musicdsp list who have practical experience of using MFCCs. Richard Dobson
|
Pages: 1 Prev: Prelim DSP design estimate Next: CompIMAGE 2010: Call for Short Communications |