HiFi-GAN

Translate mel-spectrograms into audio using Generative Adversarial Network

This is unofficial simple HiFi-GAN implementation. You can either use pre-trained weights (however this implies some artifacts in the generated audio) or train your own GAN. You can find training logs, weights and some synthesized examples here

Installation:

git clone https://github.com/ivan7022/HiFi-GAN.git

cd HiFi-GAN

pip install -r requirements.txt

To download and apply pre-trained model you simply run

python inference.py

Change some path variables or device (by default model runs on CPU) if you need.

To reproduce results / train your own GAN you should download data (by default it is LJSpeech-1.1) and run training:

bash setup_data.sh

python train.py

Do not forget to change wandb logging variables in source/config.py in this case.