Publication Details
Dereverberation and Beamforming in Far-Field Speaker Recognition
Matějka Pavel, Ing., Ph.D.
Novotný Ondřej, Ing., Ph.D.
Černocký Jan, prof. Dr. Ing. (DCGM)
Speaker recognition, microphone array,beamforming, dereverberation, audio retransmission
This paper deals with far-field speaker recognition. On a corpusof NIST SRE 2010 data retransmitted in a real roomwith multiple microphones, we first demonstrate how roomacoustics cause significant degradation of state-of-the-art ivectorbased speaker recognition system. We then investigateseveral techniques to improve the performances ranging fromprobabilistic linear discriminant analysis (PLDA) re-training,through dereverberation, to beamforming. We found thatweighted prediction error (WPE) based dereverberation combinedwith generalized eigenvalue beamformer with powerspectraldensity (PSD) weighting masks generated by neuralnetworks (NN) provides results approaching the clean closemicrophonesetup. Further improvement was obtained byre-training PLDA or the mask-generating NNs on simulatedtarget data. The work shows that a speaker recognition systemworking robustly in the far-field scenario can be developed.
@inproceedings{BUT155039,
author="Ladislav {Mošner} and Pavel {Matějka} and Ondřej {Novotný} and Jan {Černocký}",
title="Dereverberation and Beamforming in Far-Field Speaker Recognition",
booktitle="Proceedings of ICASSP 2018",
year="2018",
pages="5254--5258",
publisher="IEEE Signal Processing Society",
address="Calgary",
doi="10.1109/ICASSP.2018.8462365",
isbn="978-1-5386-4658-8",
url="https://www.fit.vut.cz/research/publication/11717/"
}