This is cloned from chrisdonahue/wavegan to learn GAN synthesize.
This is an expermental script to test original wavegan on google colaboratory.
wavegan_colab_practice1_sc09.ipynb
Following is 6 hours train result using sc09 (digits speech).
D-loss G-loss (until 3k)
There are generated wave samples in model.ckp-2995-generated-wave-sample folder.
Their sound quality is sill low, due to train time is short.
This is a conditional wavegan, by adding condition to original wavegan using drum wave as label into wav 2nd channel.
Following script is to test this conditional wavegan.
conditional_wavegan_colab_practice2_drum.ipynb
There are generated drum samples in model.ckpt-4532-generated-drum-sample folder.
This conditional result is still dissatisfaction. This condtion is probably not enough too sharp to discrimate.
Regarding to original wavegan license, please see LICENSE-wavegan.txt.
MIT