- PR https://github.com/keithito/tacotron/pull/41 adds a TensorFlow implementation of Griffin-Lim.
- Audio can be generated at 7.5x real-time on a GTX 1080Ti.
- PR https://github.com/keithito/tacotron/pull/40 makes the text processing pipeline more configurable
- This can make it easier to train on non-English data.
- See [TRAINING_DATA.md](https://github.com/keithito/tacotron/blob/master/TRAINING_DATA.md) for details.
- TensorFlow 1.3 is now required.