An implementation and demonstration of Google WaveNet.

This application learns the given sound file and the resultant model can be used to synthesize sounds.

It is a work in progress, and will be used in Cybercom #HackingGreat event in November 2016.

At the moment of writing CuDNN does not have applicable GPU acceleration functions to accelerate strided convolutions, so we have to use a trick to swap time dimensions to batch dimension and back for the normal 1-stride convolution to use GPU acceleration.

The application takes corpus.wav as the input file and test.wav as the test set file and trains the WaveNet with that corpus using different amplitude and noise distortions.

Running process.py trains a model, and saves the best model found so far into the file sound-model-best and sound-model-best.meta.

After having a trained model, you can use it to generate random sounds, by running generate.py. It saves the generated sound to Octave file sound.mat, and the probability distribution used for generation in image.mat. For generating really long samples, you might want to disable the image.mat generation by commenting out the relevant parts in generate.py.

You can draw the probability distribution and play intermediate training results in Octave using the Octave script file: plot_training.m

You can show the generated results and play the generated sound in Octave using the Octave script file: plot_generation.m