This implementation speeds up Wavenet generation by eliminating redundant convolution operations. A naive implementation of Wavenet generation is O(2^L), while ours is O(L), where L is the number of layers.
Other attempts at implementation, with current star count:
https://github.com/ibab/tensorflow-wavenet (1029 stars)
Source: https://www.reddit.com/user/kcimc | https://www.reddit.com/r/MachineLearning/comments/53ilcr/fast_wavenet_an_efficient_wavenet_generation/