I am trying to create a simple recognition system. i have recordings of several speakers saying the same word, let's say "west". I compute the MFCC feature vectors of word "west" from each speaker. If i have 10 speakers, i then have 10 matrices, each of them of dimensions N_frames * 13 (13 coeff and N_frames varies from speaker to speaker). I would now like to store a final matrix of dimension M * 13 which will serve as a reference pattern for word "west". My question is, how to "average" or combine MFCCs from different speakers? (i call this process some kind of training)
What you can try is the DTW (Dynamic Time Warping) distance (https://en.wikipedia.org/wiki/Dynamic_time_warping).
for each instance of the spoken word "west", you compute MFCC's per frame and then calculate the DTW distance between each of the pre-recorded instances of word. this gives the rough estimate of the valid DTW distance. now when a new instance of the word comes we again calculate DTW with one of the earlier instance and if the DTW distance lies roughly in the range calculated before, it is the same word.