ISCA Archive ASVSPOOF 2021
ISCA Archive ASVSPOOF 2021

CRIM's System Description for the ASVSpoof2021 Challenge

Woo Hyun Kang, Jahangir Alam, Abderrahim Fathan

In this paper, we provide description of our submitted systems to the ASVSpoof2021 Challenge logical access attack and audio deep fake tasks. The challenge provides a difficult set of trial speech samples that have been degraded through severe post-processing, including VoIP network transmission or compression. In order to detect the spoof attacks under such adversaries, we have trained multiple systems on a training set augmented with various audio codecs. We built end-to-end countermeasure systems employing residual neural networks and time delay neural networks. Furthermore, in order to analyze and employ the distributive pattern of the frame-level representations for detecting the spoof attacks, we adopted a higher order statistics pooling (HOSP) method for extracting the utterance-level embedding, which have been proven to improve the performance on the ASVSpoof2019 trial set. To exploit the complementary information learned by different model architectures, we have employed activation ensemble and fused the scores from different systems to obtain the final decision score for spoof detection. The results show that using codec augmentation, activation ensemble and HOSP technique can help the system to be more robust against trials with adversarial conditions, and further improvement could be made by performing score-level fusion among different systems.


doi: 10.21437/ASVSPOOF.2021-16

Cite as: Kang, W.H., Alam, J., Fathan, A. (2021) CRIM's System Description for the ASVSpoof2021 Challenge. Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 100-106, doi: 10.21437/ASVSPOOF.2021-16

@inproceedings{kang21b_asvspoof,
  author={Woo Hyun Kang and Jahangir Alam and Abderrahim Fathan},
  title={{CRIM's System Description for the ASVSpoof2021 Challenge}},
  year=2021,
  booktitle={Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge},
  pages={100--106},
  doi={10.21437/ASVSPOOF.2021-16}
}