This paper presents the LAboratoire de Recherche en Interaction Humain-Système (LARIHS) reassessment of the JHU's Squeeze-Excitation and Residual neTwork (ASSERT) system for logical access (LA) ASVspoof 2021 challenge. In this work, we investigate the performance of the linear filter cepstral coefficients (LFCCs) baseline low-level features and the log-linear filter banks textrograms (logLFB-textrograms) on the squeeze-excitation residual network (SENet) using more enhanced loss functions i.e., angular margin-based softmax and large margin cosine loss. In addition, we propose new thin SENet backbone versions that use the Max-Feature-Map activation function. According to the challenge progress phase results, logLFB textrograms for ASSERT with ResNet18 and MFM activation system has been chosen to be our submitted system and achieved 0.6741 and 0.3316 min-tDCF for evaluation and progress phases respectively. Moreover, the proposed ASSERT version leads to 60 % min-tDCF relative improvements when compared with the original one for the LA ASVspoof 2019 evaluation part.
Cite as: Benhafid, Z., Selouani, S.A., Yakoub, M.S., Amrouche, A. (2021) LARIHS ASSERT Reassessment for Logical Access ASVspoof 2021 Challenge. Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 94-99, doi: 10.21437/ASVSPOOF.2021-15
@inproceedings{benhafid21_asvspoof,
author={Zhor Benhafid and Sid Ahmed Selouani and Mohammed Sidi Yakoub and Abderrahmane Amrouche},
title={{LARIHS ASSERT Reassessment for Logical Access ASVspoof 2021 Challenge}},
year=2021,
booktitle={Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge},
pages={94--99},
doi={10.21437/ASVSPOOF.2021-15}
}