00mjk / multiband-wavernn

An unofficial implement of autoregressive vocoder Multiband-WaveRNN. Audio samples in https://rongjiehuang.github.io/Multiband-WaveRNN/

Multiband-WaveRNN

Pytorch implementation of MultiBand-WaveRNN model from Efficient Neural Audio Synthesis DURATION INFORMED ATTENTION NETWORK FOR MULTIMODAL SYNTHESIS

Issues

RAW mode, Unbatched generation supported. Welcome for your contribution to implement MOL mode.

Installation

Ensure you have:

• Python >= 3.6
• Pytorch 1 with CUDA

Then install the rest with pip:

pip install -r requirements.txt

How to Use

Training your own Models

Download the LJSpeech Dataset.

Edit hparams.py, point wav_path to your dataset and run:

python preprocess.py

or use preprocess.py --path to point directly to the dataset

---

Here's my recommendation on what order to run things:

1 - Train WaveRNN with:

python train_wavernn.py

2 - Generate Sentences with both models using:

python gen_wavernn.py

Speech

Mandarin

Speaker	Recording	WaveRNN	Parallel WaveGAN	FB MelGAN	SingVocoder
#1
#2
#3
#4
#5

English

Speaker	Recording	WaveRNN	Parallel WaveGAN	FB MelGAN	SingVocoder
#1
#2
#3
#4
#5

References

• DURATION INFORMED ATTENTION NETWORK FOR MULTIMODAL SYNTHESIS
• Efficient Neural Audio Synthesis

Acknowlegements

• https://github.com/fatchord/WaveRNN