After the previous steps, training and testing can be achieved through 3_train.py and 4_test.py.

The first script trains a Tacotron model on the prepared training set across the NB_EPOCHS epochs, and then saves the model in the /results folder. 

The second script allows the user to apply the previously saved model on any transcript of testing dataset. The selection of the audio to predict is done through a  variable, item_index, which should contain the index (in the testing dataset) of the wanted item. 

The estimated spectrogram is then converted to a waveform through the Griffin-Lim algorithm. The conversion function, from_spectro_to_waveform, is defined in the  /processing/proc_audio.py file.

We strongly encourage the reader to change the default settings of the code base and to play around with more advanced approaches, in order to improve the quality of the generated waveform or the convergence speed of the model.