DiffWave

DiffWave is a fast, high-quality neural vocoder and waveform synthesizer. It starts with Gaussian noise and converts it into speech via iterative refinement. The speech can be controlled by providing a conditioning signal (e.g. log-scaled Mel spectrogram). The model and architecture details are described in DiffWave: A Versatile Diffusion Model for Audio Synthesis.

What's new (2021-04-01)

• fast sampling algorithm based on v3 of the DiffWave paper

What's new (2020-10-14)

• new pretrained model trained for 1M steps
• updated audio samples with output from new model

Status (2021-04-01)

• fast inference procedure
• stable training
• high-quality synthesis
• mixed-precision training
• multi-GPU training
• command-line inference
• programmatic inference API
• PyPI package
• audio samples
• pretrained models
• unconditional waveform synthesis

Big thanks to Zhifeng Kong (lead author of DiffWave) for pointers and bug fixes.

Audio samples

22.05 kHz audio samples

Pretrained models

22.05 kHz pretrained model (31 MB, SHA256: d415d2117bb0bba3999afabdd67ed11d9e43400af26193a451d112e2560821a8)

This pre-trained model is able to synthesize speech with a real-time factor of 0.87 (smaller is faster).

Pre-trained model details

• trained on 4x 1080Ti
• default parameters
• single precision floating point (FP32)
• trained on LJSpeech dataset excluding LJ001* and LJ002*
• trained for 1000578 steps (1273 epochs)

Install

Install using pip:

pip install diffwave

or from GitHub:

git clone https://github.com/lmnt-com/diffwave.git
cd diffwave
pip install .

Training

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 22.05 kHz. If you need to change this value, edit params.py.

python -m diffwave.preprocess /path/to/dir/containing/wavs
python -m diffwave /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 (but noisy) speech by ~8k steps (~1.5h on a 2080 Ti).

Multi-GPU training

By default, this implementation uses as many GPUs in parallel as returned by torch.cuda.device_count(). You can specify which GPUs to use by setting the CUDA_DEVICES_AVAILABLE environment variable before running the training module.

Inference API

Basic usage:

from diffwave.inference import predict as diffwave_predict

model_dir = '/path/to/model/dir'
spectrogram = # get your hands on a spectrogram in [N,C,W] format
audio, sample_rate = diffwave_predict(spectrogram, model_dir, fast_sampling=True)

# audio is a GPU tensor in [N,T] format.

Inference CLI

python -m diffwave.inference --fast /path/to/model /path/to/spectrogram -o output.wav

References

• DiffWave: A Versatile Diffusion Model for Audio Synthesis
• Denoising Diffusion Probabilistic Models
• Code for Denoising Diffusion Probabilistic Models