Publication Details
Analyzing speaker verification embedding extractors and back-ends under language and channel mismatch
Stafylakis Themos
Mošner Ladislav, Ing. (DCGM)
Plchot Oldřich, Ing., Ph.D. (DCGM)
Rohdin Johan Andréas, M.Sc., Ph.D. (DCGM)
Matějka Pavel, Ing., Ph.D. (DCGM)
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
Glembek Ondřej, Ing., Ph.D.
Brummer Johan Nikolaas Langenhoven, Dr.
speaker, verification, embedding
In this paper, we analyze the behavior and performance of speaker embeddings and the back-end scoring model under domain and language mismatch. We present our findings regarding ResNet-based speaker embedding architectures and show that reduced temporal stride yields improved performance. We then consider a PLDA back-end and show how a combination of small speaker subspace, language-dependent PLDA mixture, and nuisance-attribute projection can have a drastic impact on the performance of the system. Besides, we present an efficient way of scoring and fusing class posterior logit vectors recently shown to perform well on speaker verification task. The experiments are performed using the NIST SRE 2021 setup.
@inproceedings{BUT179660,
author="Anna {Silnova} and Themos {Stafylakis} and Ladislav {Mošner} and Oldřich {Plchot} and Johan Andréas {Rohdin} and Pavel {Matějka} and Lukáš {Burget} and Ondřej {Glembek} and Johan Nikolaas Langenhoven {Brummer}",
title="Analyzing speaker verification embedding extractors and back-ends under language and channel mismatch",
booktitle="Proceedings of The Speaker and Language Recognition Workshop (Odyssey 2022)",
year="2022",
pages="9--16",
publisher="International Speech Communication Association",
address="Beijing",
doi="10.21437/Odyssey.2022-2",
url="https://www.isca-speech.org/archive/pdfs/odyssey_2022/silnova22_odyssey.pdf"
}