Clone or download

WaveNet vocoder

PyPI Build Status Build status DOI

The goal of the repository is to provide an implementation of the WaveNet vocoder, which can generate high quality raw speech samples conditioned on linguistic or acoustic features.

Audio samples are available at

Online TTS demo

A notebook supposed to be executed on is available:


  • Focus on local and global conditioning of WaveNet, which is essential for vocoder.
  • Mixture of logistic distributions loss / sampling
  • Various audio samples and pre-trained models
  • Fast inference by caching intermediate states in convolutions. Similar to arXiv:1611.09482

Pre-trained models

Note: This is not itself a text-to-speech (TTS) model. With a pre-trained model provided here, you can synthesize waveform given a mel spectrogram, not raw text. You will need mel-spectrogram prediction model (such as Tacotron2) to use the pre-trained models for TTS.

Note: As for the pretrained model for LJSpeech, the model was fine-tuned multiple times and trained for more than 1000k steps in total. Please refer to the issues (#1, #75, #45) to know how the model was trained.

Model URL Data Hyper params URL Git commit Steps
link LJSpeech link 2092a64 1000k~ steps
link CMU ARCTIC link b1a1076 740k steps

To use pre-trained models, first checkout the specific git commit noted above. i.e.,

git checkout ${commit_hash}

And then follows "Synthesize from a checkpoint" section in the README. Note that old version of may not accept --preset=<json> parameter and you might have to change according to the preset (json) file.

You could try for example:

# Assuming you have downloaded LJSpeech-1.1 at ~/data/LJSpeech-1.1
# pretrained model (20180510_mixture_lj_checkpoint_step000320000_ema.pth)
# hparams (20180510_mixture_lj_checkpoint_step000320000_ema.json)
git checkout 2092a64
python ljspeech ~/data/LJSpeech-1.1 ./data/ljspeech \
python --preset=20180510_mixture_lj_checkpoint_step000320000_ema.json \
  --conditional=./data/ljspeech/ljspeech-mel-00001.npy \
  20180510_mixture_lj_checkpoint_step000320000_ema.pth \

You can find a generated wav file in generated directory. Wonder how it works? then take a look at code:)


  • Python 3
  • CUDA >= 8.0
  • PyTorch >= v0.4.0
  • TensorFlow >= v1.3


The repository contains a core library (PyTorch implementation of the WaveNet) and utility scripts. All the library and its dependencies can be installed by:

git clone && cd wavenet_vocoder
pip install -e ".[train]"

If you only need the library part, then you can install it by the following command:

pip install wavenet_vocoder

Getting started

Preset parameters

There are many hyper parameters to be turned depends on data. For typical datasets, parameters known to work good (preset) are provided in the repository. See presets directory for details. Notice that


accepts --preset=<json> optional parameter, which specifies where to load preset parameters. If you are going to use preset parameters, then you must use same --preset=<json> throughout preprocessing, training and evaluation. e.g.,

python --preset=presets/cmu_arctic_8bit.json cmu_arctic ~/data/cmu_arctic
python --preset=presets/cmu_arctic_8bit.json --data-root=./data/cmu_arctic

instead of

python cmu_arctic ~/data/cmu_arctic
# warning! this may use different hyper parameters used at preprocessing stage
python --preset=presets/cmu_arctic_8bit.json --data-root=./data/cmu_arctic

0. Download dataset

1. Preprocessing


python ${dataset_name} ${dataset_path} ${out_dir} --preset=<json>

Supported ${dataset_name}s for now are

  • cmu_arctic (multi-speaker)
  • ljspeech (single speaker)

Assuming you use preset parameters known to work good for CMU ARCTIC dataset and have data in ~/data/cmu_arctic, then you can preprocess data by:

python cmu_arctic ~/data/cmu_arctic ./data/cmu_arctic --preset=presets/cmu_arctic_8bit.json

When this is done, you will see time-aligned extracted features (pairs of audio and mel-spectrogram) in ./data/cmu_arctic.

2. Training

Note: for multi gpu training, you have better ensure that batch_size % num_gpu == 0


python --data-root=${data-root} --preset=<json> --hparams="parameters you want to override"

Important options:

  • --speaker-id=<n>: (Multi-speaker dataset only) it specifies which speaker of data we use for training. If this is not specified, all training data are used. This should only be specified when you are dealing with a multi-speaker dataset. For example, if you are trying to build a speaker-dependent WaveNet vocoder for speaker awb of CMU ARCTIC, then you have to specify --speaker-id=0. Speaker ID is automatically assigned as follows:
In [1]: from nnmnkwii.datasets import cmu_arctic

In [2]: [(i, s) for (i,s) in enumerate(cmu_arctic.available_speakers)]

[(0, 'awb'),
 (1, 'bdl'),
 (2, 'clb'),
 (3, 'jmk'),
 (4, 'ksp'),
 (5, 'rms'),
 (6, 'slt')]

Training un-conditional WaveNet

python --data-root=./data/cmu_arctic/

You have to disable global and local conditioning by setting gin_channels and cin_channels to negative values.

Training WaveNet conditioned on mel-spectrogram

python --data-root=./data/cmu_arctic/ --speaker-id=0 \

Training WaveNet conditioned on mel-spectrogram and speaker embedding

python --data-root=./data/cmu_arctic/ \

3. Monitor with Tensorboard

Logs are dumped in ./log directory by default. You can monitor logs by tensorboard:

tensorboard --logdir=log

4. Synthesize from a checkpoint


python ${checkpoint_path} ${output_dir} --preset=<json> --hparams="parameters you want to override"

Important options:

  • --length=<n>: (Un-conditional WaveNet only) Number of time steps to generate.
  • --conditional=<path>: (Required for onditional WaveNet) Path of local conditional features (.npy). If this is specified, number of time steps to generate is determined by the size of conditional feature.


python --hparams="parameters you want to override" \
    checkpoints_awb/checkpoint_step000100000.pth \
    generated/test_awb \


Synthesize audio samples for testset


python ${checkpoint_path} ${output_dir} --data-root="data location"\
    --preset=<json> --hparams="parameters you want to override"

This script is used for generating sounds for


  • --data-root: Data root. This is required to collect testset.
  • --num-utterances: (For multi-speaker model) number of utterances to be generated per speaker. This is useful especially when testset is large and don't want to generate all utterances. For single speaker dataset, you can hit ctrl-c whenever you want to stop evaluation.


python --data-root=./data/cmu_arctic/ \
    ./checkpoints_awb/checkpoint_step000100000.pth \