Publication Details
Speaker Verification Using End-To-End Adversarial Language Adaptation
Stafylakis Themos
Silnova Anna, M.Sc., Ph.D. (DCGM)
Zeinali Hossein, Ph.D. (DCGM)
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
Plchot Oldřich, Ing., Ph.D. (DCGM)
Speaker recognition, domain adaptation
In this paper we investigate the use of adversarial domain adaptation for addressing the problem of language mismatch between speaker recognition corpora. In the context of speaker verification, adversarial domain adaptation methods aim at minimizing certain divergences between the distribution that the utterance-level features follow (i.e. speaker embeddings) when drawn from source and target domains (i.e. languages), while preserving their capacity in recognizing speakers. Neural architectures for extracting utterancelevel representations enable us to apply adversarial adaptation methods in an end-to-end fashion and train the network jointly with the standard cross-entropy loss. We examine several configurations, such as the use of (pseudo-)labels on the target domain as well as domain labels in the feature extractor, and we demonstrate the effectiveness of our method on the challenging NIST SRE16 and SRE18 benchmarks.
@inproceedings{BUT158086,
author="Johan Andréas {Rohdin} and Themos {Stafylakis} and Anna {Silnova} and Hossein {Zeinali} and Lukáš {Burget} and Oldřich {Plchot}",
title="Speaker Verification Using End-To-End Adversarial Language Adaptation",
booktitle="Proceedings of ICASSP 2019",
year="2019",
pages="6006--6010",
publisher="IEEE Signal Processing Society",
address="Brighton",
doi="10.1109/ICASSP.2019.8683616",
isbn="978-1-5386-4658-8",
url="https://ieeexplore.ieee.org/abstract/document/8683616"
}