Audio data augmentation in PyTorch. Inspired by audiomentations.
- Supports CPU and GPU (CUDA) - speed is a priority
- Supports batches of multichannel (or mono) audio
- Transforms extend
nn.Module
, so they can be integrated as a part of a pytorch neural network model - Most transforms are differentiable
- Three modes:
per_batch
,per_example
andper_channel
- Cross-platform compatibility
- Permissive MIT license
- Aiming for high test coverage
pip install torch-audiomentations
import torch
from torch_audiomentations import Compose, Gain, PolarityInversion
# Initialize augmentation callable
apply_augmentation = Compose(
transforms=[
Gain(
min_gain_in_db=-15.0,
max_gain_in_db=5.0,
p=0.5,
),
PolarityInversion(p=0.5)
]
)
torch_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Make an example tensor with white noise.
# This tensor represents 8 audio snippets with 2 channels (stereo) and 2 s of 16 kHz audio.
audio_samples = torch.rand(size=(8, 2, 32000), dtype=torch.float32, device=torch_device) - 0.5
# Apply augmentation. This varies the gain and polarity of (some of)
# the audio snippets in the batch independently.
perturbed_audio_samples = apply_augmentation(audio_samples, sample_rate=16000)
Contributors welcome!
Join the Asteroid's slack
to start discussing about torch-audiomentations
with us.
We don't want data augmentation to be a bottleneck in model training speed. Here is a comparison of the time it takes to run 1D convolution:
torch-audiomentations is in an early development stage, so the APIs are subject to change.
Every transform has mode
, p
, and p_mode
-- the parameters that decide how the augmentation is performed.
mode
decides how the randomization of the augmentation is grouped and applied.p
decides the on/off probability of applying the augmentation.p_mode
decides how the on/off of the augmentation is applied.
This visualization shows how different combinations of mode
and p_mode
would perform an augmentation.
Added in v0.5.0
Add background noise to the input audio.
Added in v0.7.0
Add colored noise to the input audio.
Added in v0.5.0
Convolve the given audio with impulse responses.
Added in v0.9.0
Apply band-pass filtering to the input audio.
Added in v0.10.0
Apply band-stop filtering to the input audio. Also known as notch filter.
Added in v0.1.0
Multiply the audio by a random amplitude factor to reduce or increase the volume. This technique can help a model become somewhat invariant to the overall gain of the input audio.
Warning: This transform can return samples outside the [-1, 1] range, which may lead to clipping or wrap distortion, depending on what you do with the audio in a later stage. See also https://en.wikipedia.org/wiki/Clipping_(audio)#Digital_clipping
Added in v0.8.0
Apply high-pass filtering to the input audio.
Added in v0.11.0
This transform returns the input unchanged. It can be used for simplifying the code in cases where data augmentation should be disabled.
Added in v0.8.0
Apply low-pass filtering to the input audio.
Added in v0.2.0
Apply a constant amount of gain, so that highest signal level present in each audio snippet in the batch becomes 0 dBFS, i.e. the loudest level allowed if all samples must be between -1 and 1.
This transform has an alternative mode (apply_to="only_too_loud_sounds") where it only applies to audio snippets that have extreme values outside the [-1, 1] range. This is useful for avoiding digital clipping in audio that is too loud, while leaving other audio untouched.
Added in v0.9.0
Pitch-shift sounds up or down without changing the tempo.
Added in v0.1.0
Flip the audio samples upside-down, reversing their polarity. In other words, multiply the waveform by -1, so negative values become positive, and vice versa. The result will sound the same compared to the original when played back in isolation. However, when mixed with other audio sources, the result may be different. This waveform inversion technique is sometimes used for audio cancellation or obtaining the difference between two waveforms. However, in the context of audio data augmentation, this transform can be useful when training phase-aware machine learning models.
Added in v0.5.0
Shift the audio forwards or backwards, with or without rollover
Added in v0.6.0
Given multichannel audio input (e.g. stereo), shuffle the channels, e.g. so left can become right and vice versa. This transform can help combat positional bias in machine learning models that input multichannel waveforms.
If the input audio is mono, this transform does nothing except emit a warning.
Added in v0.10.0
Reverse (invert) the audio along the time axis similar to random flip of an image in the visual domain. This can be relevant in the context of audio classification. It was successfully applied in the paper AudioCLIP: Extending CLIP to Image, Text and Audio
- Add new transforms:
Mix
,Padding
,RandomCrop
andSpliceOut
- Add new transform:
Identity
- Add API for processing targets alongside inputs. Some transforms experimentally support this feature already.
- Add
ObjectDict
output type as alternative totorch.Tensor
. This alternative is opt-in for now (for backwards-compatibility), but note that the old output type (torch.Tensor
) is deprecated and support for it will be removed in a future version. - Allow specifying a file path, a folder path, a list of files or a list of folders to
AddBackgroundNoise
andApplyImpulseResponse
- Require newer version of
torch-pitch-shift
to ensure support for torchaudio 0.11 inPitchShift
- Fix a bug where
BandPassFilter
didn't work on GPU
- Add support for min SNR == max SNR in
AddBackgroundNoise
- Add support for librosa 0.9.0
- Fix a bug where loaded audio snippets were sometimes resampled to an incompatible
length in
AddBackgroundNoise
- Implement
OneOf
andSomeOf
for applying one or more of a given set of transforms - Implement new transforms:
BandStopFilter
andTimeInversion
- Put
ir_paths
in transform_parameters inApplyImpulseResponse
so it is possible to inspect what impulse responses were used. This also givesfreeze_parameters()
the expected behavior.
- Fix a bug where the actual bandwidth was twice as large as expected in
BandPassFilter
. The default values have been updated accordingly. If you were previously specifyingmin_bandwidth_fraction
and/ormax_bandwidth_fraction
, you now need to double those numbers to get the same behavior as before.
- Officially mark python>=3.9 as supported
- Add parameter
compensate_for_propagation_delay
inApplyImpulseResponse
- Implement
BandPassFilter
- Implement
PitchShift
- Support for torchaudio<=0.6 has been removed
- Implement
HighPassFilter
andLowPassFilter
- Support for torchaudio<=0.6 is deprecated and will be removed in the future
- Support for pytorch<=1.6 has been removed
- Implement
AddColoredNoise
- Support for pytorch<=1.6 is deprecated and will be removed in the future
- Implement
ShuffleChannels
- Fix a bug where
AddBackgroundNoise
did not work on CUDA - Fix a bug where symlinked audio files/folders were not found when looking for audio files
- Use torch.fft.rfft instead of the torch.rfft (deprecated in pytorch 1.7) when possible. As a
bonus, the change also improves performance in
ApplyImpulseResponse
.
- Release
AddBackgroundNoise
andApplyImpulseResponse
- Implement
Shift
- Make
sample_rate
optional. Allow specifyingsample_rate
in__init__
instead offorward
. This means torchaudio transforms can be used inCompose
now.
- Remove support for 1-dimensional and 2-dimensional audio tensors. Only 3-dimensional audio tensors are supported now.
- Fix a bug where one could not use the
parameters
method of thenn.Module
subclass - Fix a bug where files with uppercase filename extension were not found
- Implement
Compose
for applying multiple transforms - Implement utility functions
from_dict
andfrom_yaml
for loading data augmentation configurations from dict, json or yaml - Officially support differentiability in most transforms
- Add support for alternative modes
per_batch
andper_channel
- Transforms now return the input unchanged when they are in eval mode
- Implement
PeakNormalization
- Expose
convolve
in the API
- Simplify API for using CUDA tensors. The device is now inferred from the input tensor.
- Initial release with
Gain
andPolarityInversion
A GPU-enabled development environment for torch-audiomentations can be created with conda:
conda env create
pytest
- Format python code with black
- Use Google-style docstrings
- Use explicit relative imports, not absolute imports
The development of torch-audiomentations is kindly backed by Nomono.
Thanks to all contributors who help improving torch-audiomentations.