$ git clone https://github.com/facebookresearch/vocoder-benchmark.git
$ cd vocoder-benchmark

$ python3 -m venv vocbench
# activate the virtualenv
$ source vocbench/bin/activate
# Upgrade pip
$ python -m pip install --upgrade pip
# Install dependences
$ pip install -e .

VOCODER_BENCHMARK=/path/to/vocoder-benchmark
export PATH=$VOCODER_BENCHMARK/bin:$PATH

$ chomd +x $VOCODER_BENCHMARK/bin/vocoder
$ vocoder --help
Usage: cli.py [OPTIONS] COMMAND [ARGS]...

  Vocoder benchmarking CLI.

Options:
  --help  Show this message and exit.

Commands:
  dataset           Dataset processing.
  diffwave          Create, train, or use diffwave models.
  parallel_wavegan  Create, train, or use parallel_wavegan models.
  wavegrad          Create, train, or use wavegrad models.
  wavenet           Create, train, or use wavenet models.
  wavernn           Create, train, or use wavernn models.

$ vocoder dataset --help # For more information on how to download/split dataset

# e.g. download and split LJ Speech
$ vocoder dataset download --dataset ljspeech --path ~/local/datasets/lj # Download and unzip dataset files
$ vocoder dataset split --dataset ljspeech --path ~/local/datasets/lj  # Create train / validation / test splits

$ vocoder [model-cmd] train --help

# e.g. train wavenet on LJ Speech dataset
$ vocoder wavenet train --path ~/local/models/wavenet --dataset ~/local/datasets/lj --config $VOCODER_BENCHMARK/config/wavenet_mulaw_normal.yaml

# MelGAN
$ vocoder parallel_wavegan train --path ~/local/models/melgan --dataset ~/local/datasets/lj --config $VOCODER_BENCHMARK/config/melgan.v1.yaml

# Parallel WaveGAN
$ vocoder parallel_wavegan train --path ~/local/models/parallel_wavegan --dataset ~/local/datasets/lj --config $VOCODER_BENCHMARK/config/parallel_wavegan.yaml

$ vocoder [model-cmd] synthesize --help
Usage: cli.py [model-cmd] synthesize [OPTIONS] INPUT_FILE OUTPUT_FILE

  Synthesize with the model.

Options:
  --path TEXT     Directory for the model  [required]
  --length TEXT   The length of the output sample in seconds
  --offset FLOAT  Offset in seconds of the sample
  --help          Show this message and exit.

$ vocoder [model-cmd] evaluate --help
Usage: cli.py [model-cmd] evaluate [OPTIONS]

  Evaluate a given vocoder.

Options:
  --path TEXT        Directory for the model  [required]
  --dataset TEXT     Name of the dataset to use  [required]
  --checkpoint TEXT  Checkpoint path (default: load latest checkpoint)
  --help             Show this message and exit.