nkululeko.augment

The nkululeko.augment module applies audio augmentations (noise, speed, pitch shifts, etc.) to diversify training data and mitigate overfitting.

Purpose

  • Increase dataset variability.

  • Enhance robustness to recording conditions.

  • Support class balance via synthetic samples.

Invocation

python -m nkululeko.augment --config examples/exp_emodb_os_svm.ini

(Augmentation options must be specified in [DATA] / dedicated augmentation sections; see ini_file.md.)

Common Augmentations

  • Additive noise

  • Time stretching

  • Pitch shifting

  • Room impulse simulation

INI Sketch

[DATA]
augment = noise,time_stretch
noise.snr = 10
time_stretch.factor = 1.05

Outputs

Augmented features integrated into feature extraction pipeline; logs indicate augmentation steps.

Tips

  • Limit augment layers to avoid excessive runtime.

  • Track original vs augmented counts for class balance verification.