xzm2004260 / wavevae

A Pytorch implementation of WaveVAE ("Parallel Neural Text-to-Speech")

WaveVAE

work in progress

Note that my implementation isn't stable yet.

A Pytorch Implementation of WaveVAE (Mel Spectrogram --> Waveform)

part of "Parallel Neural Text-to-Speech"

Requirements

PyTorch 0.4.1 & python 3.6 & Librosa

Examples

Step 1. Download Dataset

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

Step 2. Preprocessing (Preparing Mel Spectrogram)

python preprocessing.py --in_dir ljspeech --out_dir DATASETS/ljspeech

Step 3. Train Model

python train.py --model_name wavevae_1 --batch_size 4 --num_gpu 2

Step 4. Synthesize

--load_step CHECKPOINT : the # of the model's global training step (also depicted in the trained weight file)

python synthesize.py --model_name wavevae_1 --load_step 10000 --num_samples 5

References

• WaveNet vocoder : https://github.com/r9y9/wavenet_vocoder
• Parallel Neural Text-to-Speech : https://arxiv.org/abs/1905.08459