FloWaveNet : A Generative Flow for Raw Audio

Unofficial tensorflow implementation of the paper "FloWaveNet : A Generative Flow for Raw Audio".

Requirements

• Python 3.5
• tensorflow 1.12
• Librosa

How to use

1. Download the LJ-Speech dataset and unpack it:

>>> tar -xvf LJSpeech-1.1.tar.bz2

2. Preprocess dataset using the following command:

>>> python3 preprocessing.py --in_dir=LJSpeech-1.1 --out_dir=training_data

3. Run training:

>>> python3 train.py

Features

• Implemented Multig-gpu training
• Added Global condition features
• Mixed precision training

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.

Todo list

• [ ] Learning rate and batch size tuning for efficient multi-GPU training

Reference

• Official pytorch implementation: https://github.com/ksw0306/FloWaveNet