Agha Ali Raza

Publications/Efficient Machine Learning: Data Selection, Pruning, Active & Federated Learning

Conference · 2023

Self-Supervised Dataset Pruning for Efficient Training in Audio Anti-spoofing

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza

Interspeech 2023, Dublin, Ireland · CORE A

Summary

Training detectors for spoofed audio is increasingly expensive. Pruning the training set to its most useful examples helps, but existing methods need labels and a preliminary round of training just to score the examples. We propose a self-supervised alternative that finds important examples by clustering audio embeddings, with no scoring step. On the ASVSpoof 2019 dataset with two anti-spoofing models, it outperforms four other pruning metrics while being 91% more computationally efficient. Differences in how certain attacks are distributed help explain why.

Cite this paper

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza. Self-Supervised Dataset Pruning for Efficient Training in Audio Anti-spoofing. Interspeech 2023, Dublin, Ireland.

@inproceedings{azeemi2023self,
  author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{Self-Supervised Dataset Pruning for Efficient Training in Audio Anti-spoofing}},
  booktitle = {Interspeech 2023, Dublin, Ireland},
  year = {2023},
  url = {https://aghaaliraza.com/publications/2023--INTERSPEECH--Self-Supervised_Dataset_Pruning_for_Efficient_Training_in_Audio_Anti-spoofing.pdf}
}

More on this theme

All 8 works on this theme →