Sigment

An extensible data augmentation package for creating complex transformation pipelines for audio signals.

What is data augmentation?

Data augmentation is the creation of artificial data from original data by typically applying a transformation, or multiple transformations, to the original data. It is a common method for improving the versatility of machine learning models, in addition to providing more training examples for datasets of limited size.

In image data for example, it is common to use horizontal and vertical flipping, random cropping, zooming and additive noise for augmentation. In audio, we can use other transformations such as pitch shifting, time stretching or fading the signal in or out. Some image augmentation methods such as additive noise can also be transferred over to audio data.

Supported augmentation methods

Sigment currently provides the following augmentation methods for both mono and stereo signals. More information about each can be found in the documentation:

• Additive Gaussian White Noise
• Time Stretching and Pitch Shifting
• Edge Cropping and Random Cropping
• Linear Fading In/Out
• Normalization, Pre-Emphasis and Loudest Section Extraction
• Median Filtering
• Clipping Distortion and Reverb

It is also possible to design custom augmentation methods using a simple Transform base class provided by Sigment.

Example

Suppose we have the following stereo signal from audio.wav:

We can apply a pipeline of transformations to the signal to produce multiple augmented copies of it:

from librosa import load
from sigment import *

# Load the stereo WAV audio file
X, sr = load('audio.wav', mono=False)

# Create a complex augmentation pipeline
transform = Pipeline([
    GaussianWhiteNoise(scale=(0.001, 0.0075), p=0.65),
    ExtractLoudestSection(duration=(0.85, 0.95)),
    OneOf([
        RandomCrop(crop_size=(0.01, 0.04), n_crops=(2, 5)),
        SomeOf([
            EdgeCrop('start', crop_size=(0.05, 0.1)),
            EdgeCrop('end', crop_size=(0.05, 0.1))
        ], n=(1, 2))
    ]),
    Sometimes([
        SomeOf([
            LinearFade('in', fade_size=(0.1, 0.2)),
            LinearFade('out', fade_size=(0.1, 0.2))
        ], n=(1, 2))
    ], p=0.5),
    TimeStretch(rate=(0.8, 1.2)),
    PitchShift(n_steps=(-0.25, 0.25)),
    MedianFilter(window_size=(5, 10), p=0.5)
])

# Generate 25 augmentations of the signal X
transform.generate(X, n=25, sr=sr)

Installation

To install Sigment from PyPI, you can use pip:

pip install sigment

Components

Sigment provides two main components that can be used to construct augmentation pipelines:

• Transforms (sigment.transforms): Used to apply a specific type of transformation to the audio data.

• Quantifiers (sigment.quantifiers): Used to specify rules for how a sequence of transformations or nested quantifiers should be applied to augment the audio data.

Read the documentation and example notebooks for more information about the usage of both.

Acknowledgements

Sigment offers a familiar interface for transformations, taking inspiration from some other well-written augmentation libraries. Without the following libraries, the capabilities of Sigment would be very limited:

• aleju/imgaug: Image augmentation for machine learning experiments.
• makcedward/nlpaug: Data augmentation for NLP
• albumentations-team/albumentations: fast image augmentation library and easy to use wrapper around other libraries
• iver56/audiomentations: A Python library for audio data augmentation. Inspired by albumentations. Useful for machine learning.

Contributors

All contributions to this repository are greatly appreciated. Contribution guidelines can be found here.

Edwin Onuonga ✉️ 🌍

Sigment © 2019-2021, Edwin Onuonga - Released under the MIT License.
Authored and maintained by Edwin Onuonga.