Tensorflow implementation of "FloWaveNet: A Generative Flow for Raw Audio"
Unofficial tensorflow implementation of the paper "FloWaveNet : A Generative Flow for Raw Audio".
>>> tar -xvf LJSpeech-1.1.tar.bz2
>>> python3 preprocessing.py --in_dir=LJSpeech-1.1 --out_dir=training_data
>>> python3 train.py
With mixed precision training (enabled by default) the model can be trained for 7.5 days on a single GPU with 11Gb RAM. To use float32 training set dtype=tf.float32
and scale=1.
in hparams.py
.
Several examples of synthesis can be found here.