WaveNet-like vocoder models
Basic implementations of WaveNet and modified FFTNet in PyTorch. The project structre is brought from pytorch-template.
Requirements
- NumPy
- SciPy
- PyTorch >= 0.4.1
- tqdm
- librosa
Quick Start
The code in this repo by default will train a WaveNet (or FFTNet) using 80-dimension mel-spectrogram with linear interpolation.
Preprocess
Use preprocess.py to convert your wave files into mel-spectrograms.
python preprocess.py wave/files/folder -c config.json --out data
The preprocessed data will be stored in ./data.
You can change the configurations of "feature" in the .json file.
Train
python train.py -c config.json
Test
Use preprocess.py to convert a single wave file into mel-spectrogram feature.
python preprocess.py example.wav -c config.json --out test
The result is stored in test.npz.
Then use the latest checkpoint file in the ./saved folder to decoded test.npz back to waveform.
The generating process will run on gpu if you add --cuda.
python test.py test.npz outfile.wav -r saved/your-model-name/XXXX_XXXXXX/checkpoint-stepXXXXX.pth --cuda
That's it. Other instructions and advanced usage can be found in pytorch-template, I didn't change too much of the whole structure.
Customization
I add a new folder feature which is different from pytorch-template.
To use other feature like mfcc instead of mel-spectrogram, you can add your own function in ./feature/features.py with similar arguments style of get_logmel().
Other customization method can be found in pytorch-template.
Fast inference
In test.py I implement fast-wavenet generation process in a very naive way. Use fast_inference.py you can get a huge speed up (CPU only).
The speed is around 1500 samples/s on FFTNet and 300 samples/s on WaveNet.