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 and per_channel
• Cross-platform compatibility
• Permissive MIT license
• Aiming for high test coverage

Setup

pip install torch-audiomentations

Usage example

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)

Known issues

• Target data processing is still in an experimental state (#3). Workaround: Use freeze_parameters and unfreeze_parameters for now if the target data is audio with the same shape as the input.
• Using torch-audiomentations in a multiprocessing context can lead to memory leaks (#132). Workaround: If using torch-audiomentations in a multiprocessing context, it'll probably work better to run the transforms on CPU.
• Multi-GPU / DDP is not officially supported (#136). The author does not have a multi-GPU setup to test & fix this. Get in touch if you want to donate some hardware for this. Workaround: Run the transforms on single GPU instead.
• PitchShift does not support small pitch shifts, especially for low sample rates (#151). Workaround: If you need small pitch shifts applied to low sample rates, use PitchShift in audiomentations or torch-pitch-shift directly without the function for calculating efficient pitch-shift targets.

Contribute

Contributors welcome! Join the Asteroid's slack to start discussing about torch-audiomentations with us.

Motivation: Speed

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:

Note: Not all transforms have a speedup this impressive compared to CPU. In general, running audio data augmentation on GPU is not always the best option. For more info, see this article: https://iver56.github.io/audiomentations/guides/cpu_vs_gpu/

Current state

torch-audiomentations is in an early development stage, so the APIs are subject to change.

Waveform transforms

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.

AddBackgroundNoise

Added in v0.5.0

Add background noise to the input audio.

AddColoredNoise

Added in v0.7.0

Add colored noise to the input audio.

ApplyImpulseResponse

Added in v0.5.0

Convolve the given audio with impulse responses.

BandPassFilter

Added in v0.9.0

Apply band-pass filtering to the input audio.

BandStopFilter

Added in v0.10.0

Apply band-stop filtering to the input audio. Also known as notch filter.

Gain

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

HighPassFilter

Added in v0.8.0

Apply high-pass filtering to the input audio.

Identity

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.

LowPassFilter

Added in v0.8.0

Apply low-pass filtering to the input audio.

PeakNormalization

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.

PitchShift

Added in v0.9.0

Pitch-shift sounds up or down without changing the tempo.

PolarityInversion

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.

Shift

Added in v0.5.0

Shift the audio forwards or backwards, with or without rollover

ShuffleChannels

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.

TimeInversion

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

Changelog

Unreleased

Added

• Add new transforms: Mix, Padding, RandomCrop and SpliceOut

[v0.12.0] - 2025-01-15

Removed

• Remove librosa dependency in favor of torchaudio

[v0.11.2] - 2025-01-09

Fixed

• Fix a device-related bug in transform_parameters when training on multiple GPUs
• Fix a shape-related edge case bug in AddColoredNoise
• Fix a bug where an incompatible Path data type was passed to torchaudio.info

[v0.11.1] - 2024-02-07

Changed

• Add support for constant cutoff frequency in LowPassFilter and HighPassFilter
• Add support for min_f_decay==max_f_decay in AddColoredNoise
• Bump torchaudio dependency from >=0.7.0 to >=0.9.0

Fixed

• Fix inaccurate type hints in Shift
• Remove set_backend to avoid UserWarning from torchaudio

[v0.11.0] - 2022-06-29

Added

• Add new transform: Identity
• Add API for processing targets alongside inputs. Some transforms experimentally support this feature already.

Changed

• Add ObjectDict output type as alternative to torch.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 and ApplyImpulseResponse
• Require newer version of torch-pitch-shift to ensure support for torchaudio 0.11 in PitchShift

Fixed

• Fix a bug where BandPassFilter didn't work on GPU

[v0.10.1] - 2022-03-24