Using LVCNet to design the generator of Parallel WaveGAN and the same strategy to train it, the inference speed of the new vocoder is more than 5x faster than the original vocoder without any degradation in audio quality.

Our current works [Paper] has been accepted by ICASSP2021, and our previous works were described in Melglow.

1. prepare the data, download LJSpeech dataset from https://keithito.com/LJ-Speech-Dataset/, and save it in data/LJSpeech-1.1. Then run

   python -m vocoder.preprocess --data-dir ./data/LJSpeech-1.1 --config configs/lvcgan.v1.yaml

   The mel-sepctrums are calculated and saved in the folder temp/.

2. Training LVCNet

   python -m vocoder.train --config configs/lvcgan.v1.yaml --exp-dir exps/exp.lvcgan.v1

3. Test LVCNet

   python -m vocoder.test --config configs/lvcgan.v1.yaml --exp-dir exps/exp.lvcgan.v1

4. The experimental results, including training logs, model checkpoints and synthesized audios, are stored in the folder exps/exp.lvcgan.v1/.
   Similarity, you can also use the config file configs/pwg.v1.yaml to train a Parallel WaveGAN model.

   # training
   python -m vocoder.train --config configs/pwg.v1.yaml --exp-dir exps/exp.pwg.v1
   # test
   python -m vocoder.test --config configs/pwg.v1.yaml --exp-dir exps/exp.pwg.v1

Use the tensorboard to view the experimental training process:

tensorboard --logdir exps

Audio Samples are saved in samples/, where

• samples/*_lvc.wav are generated by LVCNet,
• samples/*_pwg.wav are generated by Parallel WaveGAN,
• samples/*_real.wav are the real audio.

LVCNet: Efficient Condition-Dependent Modeling Network for Waveform Generation, https://arxiv.org/abs/2102.10815
MelGlow: Efficient Waveform Generative Network Based on Location-Variable Convolution, https://arxiv.org/abs/2012.01684
https://github.com/kan-bayashi/ParallelWaveGAN
https://github.com/lmnt-com/diffwave