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.
Cite as: Silnova, A., Stafylakis, T., Mošner, L., Plchot, O., Rohdin, J., Matĕjka, P., Burget, L., Glembek, O., Brummer, N. (2022) Analyzing Speaker Verification Embedding Extractors and Back-Ends Under Language and Channel Mismatch. Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 9-16, doi: 10.21437/Odyssey.2022-2
@inproceedings{silnova22_odyssey,
author={Anna Silnova and Themos Stafylakis and Ladislav Mošner and Oldřich Plchot and Johan Rohdin and Pavel Matĕjka and Lukáš Burget and Ondřej Glembek and Niko Brummer},
title={{Analyzing Speaker Verification Embedding Extractors and Back-Ends Under Language and Channel Mismatch}},
year=2022,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2022)},
pages={9--16},
doi={10.21437/Odyssey.2022-2}
}