FFTNet

FFTNet copied to clipboard

Unofficial Implementation of FFTNet vocode paper.

• [x] implement the model.
• [x] implement tests.
• [x] overfit on a single batch (sanity check).
• [x] linearize weights for eval time.
• [x] measure the run-time on GPU and CPU. (1 sec audio takes ~47 secs) If anyone knows additional tricks from the paper, let me know. So far I asked the authors but nobody returned.
• [ ] train on LJSpeech spectrograms.
• [ ] distill model as in Parallel WaveNet paper.