tomlepaine/fast-wavenet

Reduces Wavenet generation complexity from O(2^L) to O(L) by caching intermediate dilated convolution states in queues, eliminating redundant computations across timesteps. The approach maintains layer-specific recurrent state buffers sized to each layer's dilation factor, enabling single-sample inference without recomputing the full computational graph. Applicable to any causal dilated CNN architecture for tasks like streaming audio generation or real-time sequence modeling.

1,773 stars. No commits in the last 6 months.