ISCA Archive ASVSPOOF 2021
ISCA Archive ASVSPOOF 2021

Investigation on activation functions for robust end-to-end spoofing attack detection system

Woo Hyun Kang, Jahangir Alam, Abderrahim Fathan

The main objective of the spoofing attack detection system is to detect the artefacts caused by the spoof generation process (i.e., speech synthesis or voice conversion) given a speech sample. Since the selection of activation function can affect the ability of the end-to-end spoof detection to focus on the relevant regions of the feature map, we investigate the effects of different activation functions in the antispoofing countermeasure task. From our results, it could be found that different activation functions enable the end-to-end system to learn complementary information for spoof detection. In order to exploit the complementarity between various activation functions, we propose to adopt the activation ensemble technique within the end-to-end system, where the outputs of different activation functions are pooled together. The proposed framework was experimented on the logical access (LA) task ASVSpoof2019 dataset and outperformed the systems using a single activation function.


doi: 10.21437/ASVSPOOF.2021-13

Cite as: Kang, W.H., Alam, J., Fathan, A. (2021) Investigation on activation functions for robust end-to-end spoofing attack detection system. Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 83-88, doi: 10.21437/ASVSPOOF.2021-13

@inproceedings{kang21_asvspoof,
  author={Woo Hyun Kang and Jahangir Alam and Abderrahim Fathan},
  title={{Investigation on activation functions for robust end-to-end spoofing attack detection system}},
  year=2021,
  booktitle={Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge},
  pages={83--88},
  doi={10.21437/ASVSPOOF.2021-13}
}