A GAN architecture to generate raw audio based on a dataset. This work is based on wavegan and it is made using Tensorflow and Keras.
For now just edit the variables within the python script to match your needs.
epochs_number = 40001
model_save_interval = 1000
audio_export_interval = 400
audio_export_per_epoch = 3
audio_samplerate = 16000
TRAIN_BUF = 2048
TEST_BUF = 128
BATCH_SIZE = 256
LATENT_DIM = 128
DIMS = (2**14,1)
gen_learning_rate = 0.0001
disc_learning_rate = 0.0002
number_of_disc_layers = 22
Then, simply use python kwgan.py to begin training. GPU usage is a must. This work was tested using pitzer at https://www.osc.edu