A tensorflow implementation of DeepMind's WaveNet architecture for generating lo-fi music, or any music, really. webscraper.py will scrape YouTube and download .mp4 audio files from a given Playlist link.

You'll need to encode these files as sparse matrices before training. Use the encode_data.py script to convert the files into T X N size matrices, where T is the number of timesteps at 44khz sampling rate and N=256 is the number of channels after running a mu companding transform, quantizing, and one-hot encoding.

train.py will construct the net. Hyperparameters here include duration, the maximum dilation of the network, the number of one-hot channels (though this should probably stay at 256), etc.

WaveNet.py contains the actual model.

Note that this is only tested for python 3.7, and requires the latest versions of tensorflow, scipy, numpy, and scikit-learn.