Abstract

This paper presents the SHADOW team’s submission to the ASVSpoof 2024 challenge. We evaluated various models, including ECAPA-TDNN, ResNet34, ConvNeXt, and S4 Structured-State-Space Models. 2D convolution-based models outperformed other types, with the best Progress set results achieved using FwSE-ResNet34 with codec augmentations. In the Track 1 Eval set, this system achieved minDCF= 0.44, a 47% improvement over the challenge baseline. For Track2, we contribute a straightforward method for combining well-calibrated speaker and spoofing detection scores into a single system. This involves calculating the posterior probability for a trial being both same-speaker and bonafide. However, the significant mismatch between the Dev and Progress/Eval sets not only complicated the selection of the best systems and codecs but also impacted the Eval set calibration and score combination. Nevertheless, we achieved a-DCF= 0.397 in Track 2, a 42% improvement over the baseline.

Metadata

publication

Proc. ASVspoof 2024, 36-42, 2024

year

2024

publication date

2024

authors

Jesús Villalba, Tiantian Feng, Thomas Thebaud, Jihwan Lee, Shrikanth Narayanan, Najim Dehak

link

https://www.isca-archive.org/asvspoof_2024/villalba24_asvspoof.pdf

resource_link

https://www.isca-archive.org/asvspoof_2024/villalba24_asvspoof.pdf

conference

Proc. ASVspoof 2024

pages

36-42