0
$\begingroup$

A speech audio sample can be converted to MFCC coefficients for further analysis. I wanted to know the effect of correlated data on a CNN. I know the process of computing the MFCC coefficient, which applies Discrete Cosine Transform (DCT) to decorrelate the filter bank coefficients and yield a compressed representation of the filter banks.

I wanted to argue on the usage of filter banks vs MFCC when we want to further analyze using a CNN.

Filter banks result in highly correlated data. MFCC result in decorrelated data.

Which of the two should i choose if i plan to train a CNN model for speech classification?

$\endgroup$

2 Answers 2

2
$\begingroup$

I actually don't believe it really matters. When applying the same idea to visual problems, you'll find that the raw RGB vs a DCT compressed version lead to similar results as seen here. If there exist a negative, one may argue that the correlated data may have a slower learning rate. Nevertheless, the best way to figure out the effect is to give both a try.

$\endgroup$
1
  • $\begingroup$ I will definitely try both of the solutions I proposed above and will attach my findings here. and also, could you please help me with finding papers like the one you shared to address my problem statement? as in, could you attach the link to the index where i can search for such papers, which i can refer for the architecture and build the model on my own. @Tophat $\endgroup$ Commented Nov 6, 2018 at 7:33
0
$\begingroup$

A strong machine learning model like a CNN, will perform better on melspectrogram than MFCC. The decorrelation is primarily of interest for classical methods like Hidden Markov Models and Gaussian Mixture Models.

Additionally the melspectrogram preserves locality, so it is more image-like than if the DCT is applied. This allows using pre-trained image classification models, which tends to give very strong performance with minimal amount of work.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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