aizamaksutova/vocoder

My implementation of HiFiGAN model

HiFiGAN

Project on GAN-based model capable of generating high fidelity speech. This rep contains my implementation of HiFiGAN model and all the steps to reimplement the pipeline

How to install?

Make sure to follow this guide

git clone https://github.com/aizamaksutova/Vocoder.git
cd Vocoder
pip install -r requirements.txt

How to inference?

First, you should create a file (e.g. samples.wav) with all the phrases you want to reproduce.

Then perform these:

python3 test.py -c config.json -r checkpoint.pth -t inference_data -o output --device='cuda'

-t inference_data is a directory where your samples are stored and output is a directory where the output results will be stored. Additionaly, if you want to change the mel-spec config, then go to Vocoder/melspec

For inferencing my model exactly you should do these steps:

python3 data.py
python3 test.py -c config.json -r model_best.pth -t inference_data -o output --device='cuda'

You can look at the results in the output dir

How to train the model by yourself?

In order to train the model you would need to perform simple steps, but wait for a long time for them to actually download all the data which is a LjDataset

chmod a+x prepare_data.sh
./prepare_data.sh
python3 train.py -c config.json -k wandb_key

All the other parameters are manually stored in the config.json, but you can look up the config options in train.py in order to change everything right from terminal.

Wandb report

Here is the link to my wandb report with all the architecture explanation and wavs + graphs