aozhi / wavegan

WaveGAN: using GANs to synthesize raw audio

WaveGAN

Official TensorFlow implementation of WaveGAN (Donahue et al. 2018) (paper) (demo) (sound examples). WaveGAN is a GAN approach designed for operation on raw, time-domain audio samples. It is related to the DCGAN approach (Radford et al. 2016), a popular GAN model designed for image synthesis. WaveGAN uses one-dimensional transposed convolutions with longer filters and larger stride than DCGAN, as shown in the figure above.

Usage

Requirements

# Will likely also work with newer versions of Tensorflow
pip install tensorflow-gpu==1.4.0
pip install scipy
pip install matplotlib

Build datasets

You can download the datasets from our paper bundled as TFRecords ...

• Speech Commands Zero through Nine (SC09) alternate link: (raw WAV files)
• Drums
• Piano alternate link: (raw WAV files)

or build your own from directories of audio files:

python data/make_tfrecord.py \
	/my/audio/folder/trainset \
	./data/customdataset \
	--ext mp3 \
	--fs 16000 \
	--nshards 64 \
	--slice_len 1.5 \

Train WaveGAN

To begin (or resume) training

python train_wavegan.py train ./train \
	--data_dir ./data/customdataset

If your results are unsatisfactory, try adding a post-processing filter with --wavegan_genr_pp or removing phase shuffle with --wavegan_disc_phaseshuffle 0.

To run a script that will dump a preview of fixed latent vectors at each checkpoint on the CPU

export CUDA_VISIBLE_DEVICES="-1"
python train_wavegan.py preview ./train

To run a (slow) script that will calculate inception score for the SC09 dataset at each checkpoint

export CUDA_VISIBLE_DEVICES="-1"
python train_wavegan.py incept ./train

To back up checkpoints every hour (GAN training will occasionally collapse)

python backup.py ./train 60

Train SpecGAN

Compute dataset moments to use for normalization

export CUDA_VISIBLE_DEVICES="-1"
python train_specgan.py moments ./train \
	--data_dir ./data/customdataset \
	--data_moments_fp ./train/moments.pkl

To begin (or resume) training

python train_specgan.py train ./train \
	--data_dir ./data/customdataset \
	--data_moments_fp ./train/moments.pkl

To run a script that will dump a preview of fixed latent vectors at each checkpoint on the CPU

export CUDA_VISIBLE_DEVICES="-1"
python train_specgan.py preview ./train \
	--data_moments_fp ./train/moments.pkl

To run a (slow) script that will calculate inception score for the SC09 dataset at each checkpoint

export CUDA_VISIBLE_DEVICES="-1"
python train_specgan.py incept ./train \
	--data_moments_fp ./train/moments.pkl

To back up checkpoints every hour (GAN training will occasionally collapse)

python backup.py ./train 60

Generation

The training scripts for both WaveGAN and SpecGAN create simple TensorFlow MetaGraphs for generating audio waveforms, located in the training directory. A simple usage is below; see this Colab notebook for additional features.

import tensorflow as tf
from IPython.display import display, Audio

# Load the graph
tf.reset_default_graph()
saver = tf.train.import_meta_graph('infer.meta')
graph = tf.get_default_graph()
sess = tf.InteractiveSession()
saver.restore(sess, 'model.ckpt')

# Create 50 random latent vectors z
_z = (np.random.rand(50, 100) * 2.) - 1

# Synthesize G(z)
z = graph.get_tensor_by_name('z:0')
G_z = graph.get_tensor_by_name('G_z:0')
_G_z = sess.run(G_z, {z: _z})

# Play audio in notebook
display(Audio(_G_z[0], rate=16000))

Evaluation

Our paper uses Inception score to (roughly) measure model performance. If you would like to compare to our reported numbers directly, you may run this script on a directory of 50,000 WAV files with 16384 samples each.

python score.py --audio_dir wavs

To reproduce our paper results (9.18 +- 0.04) for the SC09 (download) training dataset, run

python score.py --audio_dir sc09/train  --fix_length --n 18620

Attribution

If you use this code in your research, cite via the following BibTeX:

@article{donahue2018wavegan,
  title={Synthesizing Audio with Generative Adversarial Networks},
  author={Donahue, Chris and McAuley, Julian and Puckette, Miller},
  journal={arXiv:1802.04208},
  year={2018}
}