Agha Ali Raza

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

Conference · 2022

Dataset Pruning for Resource-Constrained Spoofed Audio Detection

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza

Interspeech 2022, Incheon, Korea · CORE A

Summary

Neural detectors of spoofed audio need large training sets, which puts them out of reach in computing-limited settings. We study how much of the training data can be removed while keeping performance. Finding existing data pruning metrics either too coarse or unstable, we propose Forgetting Norm, which scores individual training examples at finer granularity. On the ASVspoof 2019 dataset with two anti-spoofing models, it delivers up to 23% relative improvement over baseline pruning heuristics, allowing a large share of the data to be removed with only a small drop in performance.

Cite this paper

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Agha Ali Raza. Dataset Pruning for Resource-Constrained Spoofed Audio Detection. Interspeech 2022, Incheon, Korea.

@inproceedings{azeemi2022dataset,
  author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{Dataset Pruning for Resource-Constrained Spoofed Audio Detection}},
  booktitle = {Interspeech 2022, Incheon, Korea},
  year = {2022},
  url = {https://aghaaliraza.com/publications/2022--INTERSPEECH--Dataset_Pruning_for_Resource-constrained_Spoofed_Audio_Detection.pdf}
}

More on this theme

All 8 works on this theme →