2

I am working on a project where I am classifying coughs of a patient as either positive or negative for a certain pulmonary illness.

What I have at the moment is multiple cough events, segmented from larger recordings. I have extracted various spectral features for each cough event and want to add all of these to one feature vector to train a Logistic Regression (LR) classifier.

The problem is that each cough event is different in length and this results in my MFCC features being different in length too, which is an issue when training a LR classifier.

So, I want to know if anyone has some fix for dealing with MFCC feature vectors of different sizes and how to get them in the correct shape to use as a training vector. I feel really stupid, but I cant find anything about this online. Surely this issue has been encountered before?

Arton Dorneles
  • 1,629
  • 14
  • 18
Renier Botha
  • 830
  • 1
  • 10
  • 19
  • Can you clarify what you're using for linear regression input? That method usually works with a fixed number of features, such that all inputs are of the same size. I sounds to me as if you're trying to feed it a time series of data rather than a set of extracted features. As you're discovering, sequences of MFCCs do not serve well as LR input. – Prune Feb 17 '16 at 18:42
  • @Arton Dorneles, with Logisitic Regression input I refer to the input feature matrix provided to the algorithm for training. Thanx, yeah I am realising that now. In order to get the input matrix in the correct form, I have decided to average out over all windows in each cough event for the 13 MFCCs that I extract. But to keep some sense of time series, I will divide each recording into 3/4 (still deciding) parts. Thus each cough recording will have 3 (or 4) x 13 MFCC coefficients. Do you foresee any issues with this method? – Renier Botha Feb 17 '16 at 20:16
  • RenierBotha, probably you wanted mention @Prune instead of me, right? Unfortunately I can't help you since I don't know nothing about MFCCs. – Arton Dorneles Feb 17 '16 at 22:44
  • Have you searched for applicable solutions? I'm a newcomer to MFCC's (I interned on FFT's long ago), but I quickly found [this](http://stackoverflow.com/questions/35253243/how-to-train-a-machine-learning-algorithm-using-mfcc-coefficient-vectors) and [this](http://stackoverflow.com/questions/28562023/simple-word-detector-using-mfcc) that deal with incipient information loss and handling clips of different lengths. If you aren't finding applicable references here, then please illustrate your inputs, and I can take a shot at the problem. – Prune Feb 17 '16 at 22:52
  • 1
    For anyone else trying comprehend this, [here](http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/) is a nice intro to MFCC's. – Prune Feb 17 '16 at 22:53

0 Answers0