Counting 3,742 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Apr. 21, 2019
Sep. 12, 2016

WaveNet implementation in Keras

Based on and

Listen to a sample ğŸŽ¶!

~~Generate your own samples:

$ KERAS_BACKEND=theano python2 predict with models/run_20160920_120916/config.json predict_seconds=1~~ EDIT: The pretrained model had to be removed from the repository as it wasn't compatible with recent changes.


Activate a new python2 virtualenv (recommended):

pip install virtualenv
mkdir ~/virtualenvs && cd ~/virtualenvs
virtualenv wavenet
source wavenet/bin/activate

Clone and install requirements.

cd ~
git clone
cd wavenet
pip install -r requirements.txt

Using the tensorflow backend is not recommended at this time, see this issue


  • Sacred is used for managing training and sampling. Take a look at the documentation for more information.

  • This implementation does not support python3 as of now.


Once the first model checkpoint is created, you can start sampling.

Run: $ KERAS_BACKEND=theano python2 predict with models/<your_run_folder>/config.json predict_seconds=1

The latest model checkpoint will be retrieved and used to sample. The sample will be streamed to [run_folder]/samples, you can start listening when the first sample is generated.

Sampling options:

  • predict_seconds: float. Number of seconds to sample.
  • sample_argmax: True or False. Always take the argmax
  • sample_temperature: None or float. Controls the sampling temperature. 1.0 for the original distribution, < 1.0 for less exploitation, > 1.0 for more exploration.
  • seed: int: Controls the seed for the sampling procedure.
  • predict_initial_input: string: Path to a wav file, for which the first fragment_length samples are used as initial input.

e.g.: $ KERAS_BACKEND=theano python2 predict with models/[run_folder]/config.json predict_seconds=1


$ KERAS_BACKEND=theano python2

Or for a smaller network (less channels per layer). $ KERAS_BACKEND=theano python2 with small


In order to use the VCTK dataset, first download the dataset by running vctk/

Training is done with: $ KERAS_BACKEND=theano python2 with vctkdata

For smaller network: $ KERAS_BACKEND=theano python2 with vctkdata small


Train with different configurations: $ KERAS_BACKEND=theano python2 with 'option=value' 'option2=value' Available options:

  batch_size = 16
  data_dir = 'data'
  data_dir_structure = 'flat'
  debug = False
  desired_sample_rate = 4410
  dilation_depth = 9
  early_stopping_patience = 20
  fragment_length = 1152
  fragment_stride = 128
  keras_verbose = 1
  learn_all_outputs = True
  nb_epoch = 1000
  nb_filters = 256
  nb_output_bins = 256
  nb_stacks = 1
  predict_initial_input = ''
  predict_seconds = 1
  predict_use_softmax_as_input = False
  random_train_batches = False
  randomize_batch_order = True
  run_dir = None
  sample_argmax = False
  sample_temperature = 1
  seed = 173213366
  test_factor = 0.1
  train_only_in_receptive_field = True
  use_bias = False
  use_skip_connections = True
  use_ulaw = True
    decay = 0.0
    epsilon = None
    lr = 0.001
    momentum = 0.9
    nesterov = True
    optimizer = 'sgd'

Using your own training data:

  • Create a new data directory with a train and test folder in it. All wave files in these folders will be used as data.
    • Caveat: Make sure your wav files are supported by e.g. don't use 24bit wav and remove meta info.
  • Run with: $ python2 with 'data_dir=your_data_dir_name'
  • Test preprocessing results with: $ python2 test_preprocess with 'data_dir=your_data_dir_name'


  • Local conditioning
  • Global conditioning
  • Training on CSTR VCTK Corpus
  • CLI option to pick a wave file for the sample generation initial input. Done: see predict_initial_input.
  • Fully randomized training batches
  • Soft targets: by convolving a gaussian kernel over the one-hot targets, the network trains faster.
  • Decaying soft targets: the stdev of the gaussian kernel should slowly decay.

Uncertainties from paper:

  • It's unclear if the model is trained to predict t+1 samples for every input sample, or only for the outputs for which which $t-receptive_field$ was in the input. Right now the code does the latter.
  • There is no mention of weight decay, batch normalization in the paper. Perhaps this is not needed given enough data?

Note on computational cost:

The Wavenet model is quite expensive to train and sample from. We can however trade computation cost with accuracy and fidility by lowering the sampling rate, amount of stacks and the amount of channels per layer.

For a downsized model (4000hz vs 16000 sampling rate, 16 filters v/s 256, 2 stacks vs ??):

  • A Tesla K80 needs around ~4 minutes to generate one second of audio.
  • A recent macbook pro needs around ~15 minutes. Deepmind has reported that generating one second of audio with their model takes about 90 minutes.


This is a re-implementation of the model described in the WaveNet paper by Google Deepmind. This repository is not associated with Google Deepmind.