Utilizing VOiCES dataset for multichannel speaker verification with beamforming

Utilizing VOiCES dataset for multichannel speaker verification with beamforming

Paper in proceedings outside WoS and Scopus

VOiCES from a Distance Challenge 2019 aimed at the evaluationof speaker verification (SV) systems using single-channeltrials based on the Voices Obscured in Complex EnvironmentalSettings (VOiCES) corpus. Since it comprises recordingsof the same utterances captured simultaneously by multiple microphonesin the same environments, it is also suitable for multichannelexperiments. In this work, we design a multichanneldataset as well as development and evaluation trials forSV inspired by the VOiCES challenge. Alternatives discardingharmful microphones are presented as well. We asses theutilization of the created dataset for x-vector based SV withbeamforming as a front end. Standard fixed beamforming andNN-supported beamforming using simulated data and ideal binarymasks (IBM) are compared with another variant of NNsupportedbeamforming that is trained solely on the VOiCESdata. Lack of data revealed by experiments with VOiCESdatatrained beamformer was tackled by means of a variant ofSpecAugment applied to magnitude spectra. This approach ledto as much as 10% relative improvement in EER pushing resultscloser to those obtained by a good beamformer based on IBMs.

VOiCES from a Distance Challenge 2019 aimed at the evaluationof speaker verification (SV) systems using single-channeltrials based on the Voices Obscured in Complex EnvironmentalSettings (VOiCES) corpus. Since it comprises recordingsof the same utterances captured simultaneously by multiple microphonesin the same environments, it is also suitable for multichannelexperiments. In this work, we design a multichanneldataset as well as development and evaluation trials forSV inspired by the VOiCES challenge. Alternatives discardingharmful microphones are presented as well. We asses theutilization of the created dataset for x-vector based SV withbeamforming as a front end. Standard fixed beamforming andNN-supported beamforming using simulated data and ideal binarymasks (IBM) are compared with another variant of NNsupportedbeamforming that is trained solely on the VOiCESdata. Lack of data revealed by experiments with VOiCESdatatrained beamformer was tackled by means of a variant ofSpecAugment applied to magnitude spectra. This approach ledto as much as 10% relative improvement in EER pushing resultscloser to those obtained by a good beamformer based on IBMs.

multichannel speaker verification, application-aware beamforming

multichannel speaker verification, application-aware beamforming

MOŠNER, L.; PLCHOT, O.; ROHDIN, J.; ČERNOCKÝ, J.

2021

01.11.2020

International Speech Communication Association

Tokyo

Proceedings of Odyssey 2020 The Speaker and Language Recognition Workshop

2312-2846

Proceedings of Odyssey: The Speaker and Language Recognition Workshop Odyssey 2014, Joensuu, Finland

2020

11

Republic of Finland

187

193

7

https://www.isca-speech.org/archive/Odyssey_2020/abstracts/80.html

mosner_odyssey2020_80