Reproduce “Data Augmentation for Instrument Classification Robust to Audio Effects” on OpenResearch

About this paper

António Ramires, Xavier Serra

Researchers at Universitat Pompeu Fabra investigated applying common audio effects as data augmentation to enhance instrument classification robustness for sounds encountered in electronic music production. Their experiments demonstrated that while baseline models' accuracy substantially degraded on processed audio, augmenting training data with heavy distortion improved classification accuracy for distorted sounds by about 4%.

Key insights

• Models trained on unprocessed audio experience substantial degradation in classification accuracy when evaluated on sounds processed with common audio effects, with heavy distortion causing the largest drop (from 0.7378 to 0.3145 accuracy).
• Data augmentation with heavy distortion can specifically improve the model's robustness for classifying distorted sounds, suggesting effect-specific benefits for certain transformations.
• For most other audio effects, augmenting the training data with that effect did not consistently improve (and sometimes slightly decreased) accuracy on test sets processed with the same effect, indicating limitations in generalization or model capacity.

Summary by alphaXiv.