WaveNet implementation in Keras

Based on https://deepmind.com/blog/wavenet-generative-model-raw-audio/ and https://arxiv.org/pdf/1609.03499.pdf.

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

Listen to a sample 🎶!

Generate samples:

$ python wavenet.py predict with models/run_2016-09-14_11:32:09/config.json predict_seconds=1

Installation:

pip install -r requirements.txt

Note: this installs a modified version of Keras and the dev version of Theano.

Sampling:

Once the first model checkpoint is created, you can start sampling. A pretrained model is included, so sample away! (Trained on the chopin dataset from http://iwk.mdw.ac.at/goebl/mp3.html)

Run: $ python wavenet.py predict with models/run_2016-09-14_11:32:09/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. 0.01 seems to be a good value.
seed: int: Controls the seed for the sampling procedure. e.g.: $ python wavenet.py predict with models/[run_folder]/config.json predict_seconds=1 sampling_temperature=0.1

Training:

$ python wavenet.py

Or for a smaller network (less channels per layer). $ python wavenet.py with small

Options:

Train with different configurations: $ python wavenet.py with 'option=value' 'option2=value' Available options:

 batch_size = 64
  data_dir = 'data'
  debug = False
  desired_sample_rate = 4410
  dilation_depth = 9
  early_stopping_patience = 20
  fragment_length = 1024
  fragment_stride = 2045
  keras_verbose = 1
  learn_all_outputs = True
  nb_epoch = 1000
  nb_filters = 256
  nb_output_bins = 256
  nb_stacks = 1
  run_dir = None
  seed = 3004083
  train_only_in_receptive_field = True
  use_bias = False
  use_skip_connections = True
  use_ulaw = True
  optimizer:
    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 scipy.io.wavefile.read(): e.g. don't use 24bit wav and remove meta info.
Run with: $ python wavenet.py 'data_dir=your_data_dir_name'
Test preprocessing results with: $ python wavenet.py test_preprocess with 'data_dir=your_data_dir_name'

Todo:

Local conditioning
Global conditioning
Training on CSTR VCTK Corpus
CLI option to pick a wave file for the sample generation initial input.

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.