vocoders

Experiments for vocoder.

Model list

01_hifigan : HiFiGAN
02_hifigan_amp : HiFiGAN with AntiAliasActivation
03_bigvgan : BigVGAN-base
04_vocos : Vocos
05_bigvgan_f0 : BigVGAN-base with NSF module
06_ms_istft_bigvgan_f0 : BigVGAN-base with NSF module, iSTFT and Learnable PQMF
07_ms_bigvgan_f0 : BigVGAN-base with NSF module and Learnable PQMF
08_hifigan_f0 : HiFiGAN with NSF module
09_ms_hifigan_f0 : HiFiGAN with NSF module and learnalble PQMF
10_san_ms_hifigan_f0 : HiFiGAN with NSF module and learnable PQMF and utilizing SAN
11_wavenext : WaveNeXt (not stable)
12_bigvgan_v2_f0 : BigVGAN-v2(same generator as v1 but the loss and discriminator are different) with NSF module
13_bigvsan_v2_f0 : BigVSAN with BigVGAN-v2 setup and NSF module

Usage

The directiory for experiment is located in exp.

0. Install Rye

Please follow the installation instructions here.

After that, set up the project by running the following command:

$ rye sync

1. Preprocessing (00_preprocess)

Split dat to train/valid
Extract F0 with Harvest

Default config is located in src/vocoders/bin/conf/path/dummy.yaml, so if you run with original dataset, please change the path config file.
Current system supports single-speaker or universal vocoder, and it is assumed that the audio files to be used are placed in the wav_dir.
Additionally, the audio files are expected to be 1 channel, 16-bit, 24kHz in default settings.

Then run the preprocess script

$ cd /path/to/vocoders/exp/00_preprocess

$ ./run.sh

2. Training

Move to the 01_traindirectory in prepared experiment directories(e.g. exp/01_hifigan) and execute the following command.

# e.g. : `01_hifigan`
$ cd /path/to/vocoders/exp/01_hifigan/01_train

$ ./run.sh

3. Synthesis

Move to the 02_syn directory and execute the following command.
The default checkpoint file is set as 01_train/out/ckpt/last.ckpt.
In the process, validation data is synthesized and evaluated.