mxx657845129 / wavernn-3

Pytorch implementation of waveRNN vocoder

Recurrent Neural Network based Neural Vocoders

PyTorch implementation of waveRNN based neural vocoder, which predicts a raw waveform from a mel-spectrogram.

Getting started

0. Download dataset

• LJSpeech (en): https://keithito.com/LJ-Speech-Dataset/

1. Preprocessing

python preprocess.py \
        --dataset_dir <Path to the dataset dir (Location where the dataset is downloaded)>\
        --out_dir <Path to the output dir (Location where processed dataset will be written)>

The preprocessing code currently supports the following datasets:

• LJSpeech (en)

2. Training

python train.py \
     --train_data_dir <Path to the dir containing the data to train the model> \
     --checkpoint_dir <Path to the dir where the training checkpoints will be saved> \
     --resume_checkpoint_path <If specified load checkpoint and resume training from that point>

3. Generation

python generate.py \
    --checkpoint_path <Path to the checkpoint to use to instantiate the model> \
    --eval_data_dir <Path to the dir containing the mel spectrograms to be synthesized> \ 
    --out_dir <Path to the dir where the generated waveforms will be saved>

Acknowledgements

The code in this repository is based on the code in the following repositories

1. mkotha/WaveRNN
2. fatchord/WaveRNN
3. bshall/UniversalVocoding

References

1. arXiv:1802.08435: Efficient Neural Audio Synthesis
2. arXiv:1811.06292v2: Towards Achieving Robust Universal Neural Vocoding