PyTorch Audio Augmentations

Audio data augmentations library for PyTorch for audio in the time-domain. The focus of this repository is to:

• Provide many audio transformations in an easy Python interface.
• Have a high test coverage.
• Easily control stochastic (sequential) audio transformations.
• Make every audio transformation differentiable with PyTorch's nn.Module.
• Optimise audio transformations for CPU and GPU.

It supports stochastic transformations as used often in self-supervised, semi-supervised learning methods. One can apply a single stochastic augmentation or create as many stochastically transformed audio examples from a single interface.

This package follows the conventions set out by torchvision and torchaudio, with audio defined as a tensor of [channel, time], or a batched representation [batch, channel, time]. Each individual augmentation can be initialized on its own, or be wrapped around a RandomApply interface which will apply the augmentation with probability p.

Usage

We can define a single or several audio augmentations, which are applied sequentially to an audio waveform.

from audio_augmentations import *

audio, sr = torchaudio.load("tests/classical.00002.wav")

num_samples = sr * 5
transforms = [
    RandomResizedCrop(n_samples=num_samples),
    RandomApply([PolarityInversion()], p=0.8),
    RandomApply([Noise(min_snr=0.001, max_snr=0.005)], p=0.3),
    RandomApply([Gain()], p=0.2),
    HighLowPass(sample_rate=sr), # this augmentation will always be applied in this aumgentation chain!
    RandomApply([Delay(sample_rate=sr)], p=0.5),
    RandomApply([PitchShift(
        n_samples=num_samples,
        sample_rate=sr
    )], p=0.4),
    RandomApply([Reverb(sample_rate=sr)], p=0.3)
]

We can also define a stochastic augmentation on multiple transformations. The following will apply both polarity inversion and white noise with a probability of 80%, a gain of 20%, and delay and reverb with a probability of 50%:

transforms = [
    RandomResizedCrop(n_samples=num_samples),
    RandomApply([PolarityInversion(), Noise(min_snr=0.001, max_snr=0.005)], p=0.8),
    RandomApply([Gain()], p=0.2),
    RandomApply([Delay(sample_rate=sr), Reverb(sample_rate=sr)], p=0.5)
]

We can return either one or many versions of the same audio example:

transform = Compose(transforms=transforms)
transformed_audio =  transform(audio)
>> transformed_audio.shape = [num_channels, num_samples]

audio = torchaudio.load("testing/classical.00002.wav")
transform = ComposeMany(transforms=transforms, num_augmented_samples=4)
transformed_audio = transform(audio)
>> transformed_audio.shape = [4, num_channels, num_samples]

Similar to the torchvision.datasets interface, an instance of the Compose or ComposeMany class can be supplied to torchaudio dataloaders that accept transform=.

Optional

Install WavAugment for reverberation / pitch shifting:

pip install git+https://github.com/facebookresearch/WavAugment

Cite

You can cite this work with the following BibTeX:

@misc{spijkervet_torchaudio_augmentations,
  doi = {10.5281/ZENODO.4748582},
  url = {https://zenodo.org/record/4748582},
  author = {Spijkervet,  Janne},
  title = {Spijkervet/torchaudio-augmentations},
  publisher = {Zenodo},
  year = {2021},
  copyright = {MIT License}
}