I'm trying to implement a simple single digit classifier using PyBrain. I have training data of many recorded examples of the digits zero to nine, and have written a script to extract MFCCs for each example. To keep things simple (for now) I'm only calculating 13 coefficients, and am sampling the audio files once every 10ms. As each of the recorded training examples are of different durations, the number of samples for the set of all "zeros", for example, changes for each, depending on how long the audio is. So for some examples of "zero" I have an MFCC training vector of shape (67,13), while for others it might be (44, 13).
The problem is that when creating a dataset in PyBrain, the dimensions of the input vectors have to be uniform and should contain the same number of rows as columns.
Is there some way to define a data set so that the input vector will have a set number of columns (the 13 MFCCs) but a variable number of rows (depending on the number of samples taken from the signal) ? Or am I doing this completely wrong?
I have searched and searched online and on Stack Overflow and cannot come up with the answer. I can follow and implement the Pybrain examples in the tutorial.
