I am executing MFCC with 100msec sliding window on a 10Khz feed, I would like to measure the energy for each Mel Frequency Filter, sounds trivial so far, I would expect the follwoing to give me exactly that: 1. A sliding window in time of e.g. 100msec 2. Applying a window function ( e.g. Hamming ) to reduce leakage 3. executing FFT 4. Applying Mel, Bark ( or whatever ) Frequency Filter bank 5. Log coefficients for the Filter Bank ~buckets~ (*) Having the above applied the bucket of the highest energy should have the highest Log coeeficient. To verify this, I have synthisized a 2400Hz signal and extracted the features using the above mentioned flow, I have expected the ~6th filter to have the highest coefficient value, suprisingly, the 1st coefficient ( 0 - 1KHz ) had the highest value, WHY ? ( I assume this is somehow related with the fact that the FFT is executed on a 100msec window while the result is projected on a freq larger than the Nyquist frequency ). Can I accurately derive the energy at a specific frequecy using MFCC features ? Any help would be appreciated.
MFCC, deriving the energy at a specific freq range
Started by ●April 12, 2011
Reply by ●April 12, 20112011-04-12
On Apr 12, 7:00�am, Nadav <nadav...@gmail.com> wrote:> I am executing MFCC with 100msec sliding window on a 10Khz feed, I > would like to measure the energy for each Mel Frequency Filter, sounds > trivial so far, I would expect the follwoing to give me exactly that: > > 1. A sliding window in time of e.g. 100msec > 2. Applying a window function ( e.g. Hamming ) to reduce leakage > 3. executing FFT > 4. Applying Mel, Bark ( or whatever ) Frequency Filter bank > 5. Log coefficients for the Filter Bank ~buckets~ > (*) Having the above applied the bucket of the highest energy should > have the highest Log coeeficient. > > To verify this, I have synthisized a 2400Hz signal and extracted the > features using the above mentioned flow, I have expected the ~6th > filter to have the highest coefficient value, suprisingly, the 1st > coefficient ( 0 - 1KHz ) had the highest value, WHY ? ( I assume this > is somehow related with the fact that the FFT is executed on a 100msec > window while the result is projected on a freq larger than the Nyquist > frequency ). > > Can I accurately derive the energy at a specific frequecy using MFCC > features ? > > Any help would be appreciated.If I understand your post correctly, ask yourself what is in the first bin (0 - 1KHz) of your FFT.
