WaveGrad is a fast, high-quality neural vocoder designed by the folks at Google Brain. The architecture is described in WaveGrad: Estimating Gradients for Waveform Generation. In short, this model takes a log-scaled Mel spectrogram and converts it to a waveform via iterative refinement.
- stable training (22 kHz, 24 kHz)
- high-quality synthesis
- mixed-precision training
- custom noise schedule (faster inference)
- programmatic API
- PyPI package
- audio samples
- pretrained models
24 kHz pretrained model (183 MB, SHA256: 65e9366da318d58d60d2c78416559351ad16971de906e53b415836c068e335f3
)
Install using pip:
pip install wavegrad
or from GitHub:
git clone https://github.com/lmnt-com/wavegrad.git
cd wavegrad
pip install .
Before you start training, you'll need to prepare a training dataset. The dataset can have any directory structure as long as the contained .wav files are 16-bit mono (e.g. LJSpeech, VCTK). By default, this implementation assumes a sample rate of 22 kHz. If you need to change this value, edit params.py.
python -m wavegrad.preprocess /path/to/dir/containing/wavs
python -m wavegrad /path/to/model/dir /path/to/dir/containing/wavs
# in another shell to monitor training progress:
tensorboard --logdir /path/to/model/dir --bind_all
You should expect to hear intelligible speech by ~20k steps (~1.5h on a 2080 Ti).
Basic usage:
from wavegrad.inference import predict as wavegrad_predict
model_dir = '/path/to/model/dir'
spectrogram = # get your hands on a spectrogram in [N,C,W] format
audio, sample_rate = wavegrad_predict(spectrogram, model_dir)
# audio is a GPU tensor in [N,T] format.
If you have a custom noise schedule (see below):
from wavegrad.inference import predict as wavegrad_predict
params = { 'noise_schedule': np.load('/path/to/noise_schedule.npy') }
model_dir = '/path/to/model/dir'
spectrogram = # get your hands on a spectrogram in [N,C,W] format
audio, sample_rate = wavegrad_predict(spectrogram, model_dir, params=params)
# `audio` is a GPU tensor in [N,T] format.
python -m wavegrad.inference /path/to/model /path/to/spectrogram -o output.wav
The default implementation uses 1000 iterations to refine the waveform, which runs slower than real-time. WaveGrad is able to achieve high-quality, faster than real-time synthesis with as few as 6 iterations without re-training the model with new hyperparameters.
To achieve this speed-up, you will need to search for a noise schedule
that works well for your dataset. This implementation provides a script to perform the search for you:
python -m wavegrad.noise_schedule /path/to/trained/model /path/to/preprocessed/validation/dataset
python -m wavegrad.inference /path/to/trained/model /path/to/spectrogram -n noise_schedule.npy -o output.wav
The default settings should give good results without spending too much time on the search. If you'd like to find a better noise schedule or use a different number of inference iterations, run the noise_schedule
script with --help
to see additional configuration options.