speechaugs 0.0.11

Waveform augmentations

speechaugs

Single-channel waveforms augmentations for speech recognition models.

---

Augmentations:

Tranforms in time domain:

• Time Stretch
• Forward Time Shift

frequency domain:

• Pitch Shift
• Vocal Tract Length Perturbation

Noise injection:

• Colored Noise (white, pink, brown, blue, violet, grey)
• Short Noises
• File Noise

And changing the waveform samples directly:

• Zero Samples
• Clipping samples
• Inversion
• Loudness Change
• Normalization

---

Colab Example You can see examples of all augmentations and listen to resulting audios on this page with Colab notebook.

---

Installation

pip install speechaugs

---

Time Stretch

Stretch a wavefom in time with randomly chosen rate. Is implemented using librosa.effects.time_stretch.

Forward Time Shift

Shift a waveform forwards in time.

Pitch Shift

Shift a pitch by n_steps semitones. Is implemented using librosa.effects.pitch_shift.

The work of PitchShift can be better illustrated on the MelSpectrograms of waveforms.

Higher pitch (+9 semitones):

Lower pitch (-5 semitones)

Vocal Tract Length Perturbation

Change vocal tract length. Effect is very similar to Pitch Shift but speech sounds more natural.

Colored Noise

Add noise of different color to a waveform. Color of noise depends on the spectral density of the noise. You can go to wiki page for more information.

This class is implemented using colorednoise package. The color of noise is randomly choosen.

White Noise

Brown Noise

Short Noises

Add several short noises (of same color) to different parts of a waveform.

File Noise

Add noise from randomly chosen file from specified folder. Works with "sox_io" torchaudio backend. To change backend you can run:

torchaudio.set_audio_backend('sox_io')

Zero Samples

Set some percentage of samples to zero.

Clipping Samples

Clip some percentage of samples from a waveform.

Inversion

Change sign of waveform samples.

Loudness Change

Change loudness of intervals of a waveform. For example, in the figure below initial waveform was splitted into 3 intervals and samples from each of them were multiplied by different random factors.

Normalization

Normalize a waveform with choosen method ("minmax", "max" or "meanstd")

---

Usage example (with default parameters)

Import:

import speechaugs

Other libs:

import torch, torchaudio
import albumentations as A

Usage:

ex_waveform, sr = torchaudio.load('audio_filename')
noiseroot = 'path_to_noise_folder'

transforms = A.Compose([
    speechaugs.ForwardTimeShift(p=0.5),
    A.OneOf([speechaugs.Inversion(p=0.5), speechaugs.LoudnessChange(p=0.5)], p=0.5),
    A.OneOf([speechaugs.ZeroSamples(p=0.5), speechaugs.ClippingSamples(p=0.5)], p=0.5),
    A.OneOf([speechaugs.TimeStretchLibrosa(p=0.5), speechaugs.PitchShiftLibrosa(p=0.5)], p=0.5),
    A.OneOf([speechaugs.ColoredNoise(p=0.3), speechaugs.ShortNoises(p=0.3), speechaugs.FileNoise(noiseroot, p=0.3)], p=0.5),
], p=1.0)

augmented = transforms(waveform=ex_waveform)['waveform']