/wavenet_vocoder

WaveNet vocoder

Primary LanguagePythonOtherNOASSERTION

WaveNet vocoder

PyPI Build Status Build status DOI

NOTE: This is the development version. If you need a stable version, please checkout the v0.1.1.

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 https://r9y9.github.io/wavenet_vocoder/.

News

Online TTS demo

A notebook supposed to be executed on https://colab.research.google.com is available:

Highlights

  • Focus on local and global conditioning of WaveNet, which is essential for vocoder.
  • 16-bit raw audio modeling by mixture distributions: mixture of logistics (MoL), mixture of Gaussians, and single Gaussian distributions are supported.
  • Various audio samples and pre-trained models
  • Fast inference by caching intermediate states in convolutions. Similar to arXiv:1611.09482
  • Integration with ESPNet (https://github.com/espnet/espnet)

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 synthesis.py may not accept --preset=<json> parameter and you might have to change hparams.py 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 preprocess.py ljspeech ~/data/LJSpeech-1.1 ./data/ljspeech \
  --preset=20180510_mixture_lj_checkpoint_step000320000_ema.json
python synthesis.py --preset=20180510_mixture_lj_checkpoint_step000320000_ema.json \
  --conditional=./data/ljspeech/ljspeech-mel-00001.npy \
  20180510_mixture_lj_checkpoint_step000320000_ema.pth \
  generated

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

Repository structure

The repository consists of 1) pytorch library, 2) command line tools, and 3) ESPnet-style recipes. The first one is a pytorch library to provide WavaNet functionality. The second one is a set of tools to run WaveNet training/inference, data processing, etc. The last one is the reproducible recipes combining the WaveNet library and utility tools. Please take a look at them depending on your purpose. If you want to build your WaveNet on your dataset (I guess this is the most likely case), the recipe is the way for you.

Requirements

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

Installation

git clone https://github.com/r9y9/wavenet_vocoder && cd wavenet_vocoder
pip install -e .

If you only need the library part, you can install it from pypi:

pip install wavenet_vocoder

Getting started

Kaldi-style recipes

The repository provides Kaldi-style recipes to make experiments reproducible and easily manageable. Available recipes are as follows:

  • mulaw256: WaveNet that uses categorical output distribution. The input is 8-bit mulaw quantized waveform.
  • mol: Mixture of Logistics (MoL) WaveNet. The input is 16-bit raw audio.
  • gaussian: Single-Gaussian WaveNet (a.k.a. teacher WaveNet of ClariNet). The input is 16-bit raw audio.

All the recipe has run.sh, which specifies all the steps to perform WaveNet training/inference including data preprocessing. Please see run.sh in egs directory for details.

NOTICE: Global conditioning for multi-speaker WaveNet is not supported in the above recipes (it shouldn't be difficult to implement though). Please check v0.1.12 for the feature, or if you really need the feature, please raise an issue.

Apply recipe to your own dataset

The recipes are designed to be generic so that one can use them for any dataset. To apply recipes to your own dataset, you'd need to put all the wav files in a single flat directory. i.e.,

> tree -L 1 ~/data/LJSpeech-1.1/wavs/ | head
/Users/ryuichi/data/LJSpeech-1.1/wavs/
├── LJ001-0001.wav
├── LJ001-0002.wav
├── LJ001-0003.wav
├── LJ001-0004.wav
├── LJ001-0005.wav
├── LJ001-0006.wav
├── LJ001-0007.wav
├── LJ001-0008.wav
├── LJ001-0009.wav

That's it! The last step is to modify db_root in run.sh or give db_root as the command line argment for run.sh.

./run.sh --stage 0 --stop-stage 0 --db-root ~/data/LJSpeech-1.1/wavs/

Step-by-step

A recipe typically consists of multiple steps. It is strongly recommended to run the recipe step-by-step to understand how it works for the first time. To do so, specify stage and stop_stage as follows:

./run.sh --stage 0 --stop-stage 0
./run.sh --stage 1 --stop-stage 1
./run.sh --stage 2 --stop-stage 2

In typical situations, you'd need to specify CUDA devices explciitly expecially for training step.

CUDA_VISIBLE_DEVICES="0,1" ./run.sh --stage 2 --stop-stage 2

Docs for command line tools

Command line tools are writtern with docopt. See each docstring for the basic usages.

tojson.py

Dump hyperparameters to a json file.

Usage:

python tojson.py --hparams="parameters you want to override" <output_json_path>

preprocess.py

Usage:

python preprocess.py wavallin ${dataset_path} ${out_dir} --preset=<json>

train.py

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

Usage:

python train.py --dump-root=${dump-root} --preset=<json>\
  --hparams="parameters you want to override"

evaluate.py

Given a directoy that contains local conditioning features, synthesize waveforms for them.

Usage:

python evaluate.py ${dump_root} ${checkpoint} ${output_dir} --dump-root="data location"\
    --preset=<json> --hparams="parameters you want to override"

Options:

  • --num-utterances=<N>: Number of utterances to be generated. If not specified, generate all uttereances. This is useful for debugging.

synthesis.py

NOTICE: This is probably not working now. Please use evaluate.py instead.

Synthesize waveform give a conditioning feature.

Usage:

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

Important options:

  • --conditional=<path>: (Required for conditional 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.

Training scenarios

Training un-conditional WaveNet

NOTICE: This is probably not working now. Please check v0.1.1 for the working version.

python train.py --dump-root=./data/cmu_arctic/
    --hparams="cin_channels=-1,gin_channels=-1"

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 train.py --dump-root=./data/cmu_arctic/ --speaker-id=0 \
    --hparams="cin_channels=80,gin_channels=-1"

Training WaveNet conditioned on mel-spectrogram and speaker embedding

NOTICE: This is probably not working now. Please check v0.1.1 for the working version.

python train.py --dump-root=./data/cmu_arctic/ \
    --hparams="cin_channels=80,gin_channels=16,n_speakers=7"

Misc

Monitor with Tensorboard

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

tensorboard --logdir=log

List of papers that used the repository

Thank you very much!! If you find a new one, please submit a PR.

References

test