Publication Details

ABC SYSTEM DESCRIPTION FOR NIST SRE 2024

ALAM, J.; BARAHONA QUIRÓS, S.; BOBOŠ, D.; BURGET, L.; CUMANI, S.; DAHMANE, M.; HAN, J.; HLAVÁČEK, M.; KODOVSKÝ, M.; LANDINI, F.; MOŠNER, L.; PÁLKA, P.; PAVLÍČEK, T.; PENG, J.; PLCHOT, O.; RAJASEKHAR, P.; ROHDIN, J.; SILNOVA, A.; STAFYLAKIS, T.; ZHANG, L. ABC SYSTEM DESCRIPTION FOR NIST SRE 2024. Proceedings of NIST SRE 2024. San Juan: National Institute of Standards and Technology, 2024. p. 1-9.
Czech title
Popis ABC systému pro NIST SRE 2024 evaluace
Type
conference paper
Language
English
Authors
Alam Jahangir
BARAHONA QUIRÓS, S.
Boboš Dominik, Ing.
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
Cumani Sandro, Ph.D.
DAHMANE, M.
Han Jiangyu (DCGM)
HLAVÁČEK, M.
KODOVSKÝ, M.
Landini Federico Nicolás, Ph.D. (RG SPEECH)
Mošner Ladislav, Ing. (DCGM)
Pálka Petr, Bc. (DCGM)
Pavlíček Tomáš, Ing.
Peng Junyi (DCGM)
Plchot Oldřich, Ing., Ph.D. (DCGM)
RAJASEKHAR, P.
Rohdin Johan Andréas, M.Sc., Ph.D. (DCGM)
Silnova Anna, M.Sc., Ph.D. (DCGM)
Stafylakis Themos
Zhang Lin, Ph.D. (DCGM)
URL
Keywords

NIST, speaker, recognition, evaluation

Abstract

This paper presents the ABC team's submission to the NIST SRE 2024 evaluation,
a collaboration among BUT, Polito, Phonexia, Omilia, UAM, and CRIM. Our team
participated in all evaluation tracks (audio-only, visual-only, and audio-visual)
under both fixed and open conditions. We developed a variety of frontends, back-
ends, and strategies for calibration and fusion to optimize system performance.
The fixed and open conditions share some solutions. In the audio-only systems, we
employed ResNet variants and the newly introduced ReDimNet model as frontends for
embedding extraction. Then, we explored various backends including cosine
scoring, Prob- abilistic Linear Discriminant Analysis, and Pairwise Support Vec-
tor Machine. For the visual-only systems, we adopted the Insight- face framework,
utilized ResNet100 and MagFace pre-trained on the MS1MV2 dataset. Cosine scoring
under various strategies were ap- plied, with logistic regression used for both
calibration and fusion. Finally, scores from audio-only and visual-only systems
were fused using logistic regression for submission to the audio-visual track.
Building on the fixed condition, the open condition included en- hancements such
as larger ResNet models, additional training data from the VoxBlink2 dataset, and
the pre-trained XLS-R foundation model

Published
2024
Pages
1–9
Proceedings
Proceedings of NIST SRE 2024
Publisher
National Institute of Standards and Technology
Place
San Juan
BibTeX 
@inproceedings{BUT193961,
  author="ALAM, J. and BARAHONA QUIRÓS, S. and BOBOŠ, D. and BURGET, L. and CUMANI, S. and DAHMANE, M. and HAN, J. and HLAVÁČEK, M. and KODOVSKÝ, M. and LANDINI, F. and MOŠNER, L. and PÁLKA, P. and PAVLÍČEK, T. and PENG, J. and PLCHOT, O. and RAJASEKHAR, P. and ROHDIN, J. and SILNOVA, A. and STAFYLAKIS, T. and ZHANG, L.",
  title="ABC SYSTEM DESCRIPTION FOR NIST SRE 2024",
  booktitle="Proceedings of NIST SRE 2024",
  year="2024",
  pages="1--9",
  publisher="National Institute of Standards and Technology",
  address="San Juan",
  url="https://www.fit.vut.cz/research/publication/13341/"
}
Files
Back to top