To generate the noisy sound files, put the tidigits directory into the generate_data/ folder, and from generate_data/ run make. You will need sox to do that ! (apt-get install sox, yum install sox, and so on...)

The dataset is composed of the STFT of the clean / noise / noisy signal. Sound processing is done in the source_separation/wav_fft.py file, and the dataset is built in dataIO.py.

To det the train and test data, from the source_separation directory just do:

from preprocessing import dataIO

Xtrain, Ytrain = dataIO.train_set()
Xtest, Ytest   = dataIO.test_set()

Note that the train_set() function returns one big 2D array of all the STFT from all the files concatenated.

But the test_set() returns a list, for which each element contains the STFT for one WAV file. This enables to test the network only on a few files.