PyTorch implementation of MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis.

This implementation includes distributed support and uses the LJSpeech dataset.

1. NVIDIA GPU + CUDA cuDNN
2. Python >= 3.5
3. PyTorch == 1.2
4. Clone this repository.
5. Install python requirements. Please refer requirements.txt

1. Download and extract the LJ Speech dataset
2. Move all wav files to data/LJSpeech-1.1/wavs

1. python train.py --config=config.json --cps=cp_melgan

Default checkpoint directory is cp_melgan

Tensorboard logs will be saved in cp_melgan/logs

1. python distributed.py --config=config.json --args_str="--cps=cp_melgan"

Training code detects all GPUs and sets them automatically.

1. Current Samples (489K Steps)

   Samples

2. Sample audio can be heard on the tensorboard also.

3. Generated spectrograme can be seen on the tensorboard.

Coming soon..

I will commit a full inference code soon.

• distributed.py from NVIDIA's waveglow.

I referred to waveglow to implement audio preprocessing.