This is a TensorFlow implementation of the paper Efficient Neural Audio Synthesis found at the arxiv link here.
The first steps is to get an inefficient archetecture working. We wont be introducing custom GPU kernels or doing weight pruning until the model works well. Once we can sample the model with a degree of success, the objective is to implement the operations that make this model faster than real-time.
Contributions are very welcome. Please leave issues or get in touch with me at leonfedden (at) gmail.com