Publication Details
End-to-End DNN Based Speaker Recognition Inspired by i-Vector and PLDA
Silnova Anna, M.Sc., Ph.D. (DCGM)
Diez Sánchez Mireia, M.Sc., Ph.D. (DCGM)
Plchot Oldřich, Ing., Ph.D. (DCGM)
Matějka Pavel, Ing., Ph.D. (DCGM)
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
Speaker verification, DNN, end-to-end
Recently, several end-to-end speaker verification systems based ondeep neural networks (DNNs) have been proposed. These systemshave been proven to be competitive for text-dependent tasks as wellas for text-independent tasks with short utterances. However, fortext-independent tasks with longer utterances, end-to-end systemsare still outperformed by standard i-vector + PLDA systems. In thiswork, we develop an end-to-end speaker verification system that isinitialized to mimic an i-vector + PLDA baseline. The system isthen further trained in an end-to-end manner but regularized so thatit does not deviate too far from the initial system. In this way wemitigate overfitting which normally limits the performance of endto-end systems. The proposed system outperforms the i-vector +PLDA baseline on both long and short duration utterances.
@inproceedings{BUT155046,
author="Johan Andréas {Rohdin} and Anna {Silnova} and Mireia {Diez Sánchez} and Oldřich {Plchot} and Pavel {Matějka} and Lukáš {Burget}",
title="End-to-End DNN Based Speaker Recognition Inspired by i-Vector and PLDA",
booktitle="Proceedings of ICASSP",
year="2018",
pages="4874--4878",
publisher="IEEE Signal Processing Society",
address="Calgary",
doi="10.1109/ICASSP.2018.8461958",
isbn="978-1-5386-4658-8",
url="https://www.fit.vut.cz/research/publication/11724/"
}