Channel Adversarial Training for Speaker Verification and Diarization

Chau Luu; Peter Bell; Steve Renals

doi:10.1109/ICASSP40776.2020.9053323

Channel Adversarial Training for Speaker Verification and Diarization

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

Previous work has encouraged domain-invariance in deep speaker embedding by adversarially classifying the dataset or labelled environment to which the generated features belong. We propose a training strategy which aims to produce features that are invariant at the granularity of the recording or channel, a finer grained objective than dataset- or environment- invariance. By training an adversary to predict whether pairs of same-speaker embeddings belong to the same recording in a Siamese fashion, learned features are discouraged from utilizing channel information that may be speaker discriminative during training. Experiments for verification on VoxCeleb and diarization and verification on CALLHOME show promising improvements over a strong baseline in addition to outperforming a dataset-adversarial model. The VoxCeleb model in particular performs well, achieving a 4% relative improvement in EER over a Kaldi baseline, while using a similar architecture and less training data.

Original language	English
Title of host publication	ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Publisher	Institute of Electrical and Electronics Engineers (IEEE)
Pages	7094-7098
Number of pages	5
ISBN (Electronic)	978-1-5090-6631-5
ISBN (Print)	978-1-5090-6632-2
DOIs	https://doi.org/10.1109/ICASSP40776.2020.9053323
Publication status	Published - 14 May 2020
Event	2020 IEEE International Conference on Acoustics, Speech, and Signal Processing - Barcelona, Spain Duration: 4 May 2020 → 8 May 2020 Conference number: 45

Publication series

Name
Publisher	IEEE
ISSN (Print)	1520-6149
ISSN (Electronic)	2379-190X

Conference

Conference	2020 IEEE International Conference on Acoustics, Speech, and Signal Processing
Abbreviated title	ICASSP 2020
Country	Spain
City	Barcelona
Period	4/05/20 → 8/05/20

Keywords

Speaker verification
diarization
domain adversarial training
adversarial learning
deep neural network

Access to Document

10.1109/ICASSP40776.2020.9053323Licence: All Rights Reserved

Channel Adversarial Training_LAU_DOA24012020_AFVAccepted author manuscript, 237 KB

https://ieeexplore.ieee.org/document/9053323Licence: All Rights Reserved

Fingerprint Dive into the research topics of 'Channel Adversarial Training for Speaker Verification and Diarization'. Together they form a unique fingerprint.

Cite this

@inproceedings{cac3db854f3042e0b3a3e67117304602,

title = "Channel Adversarial Training for Speaker Verification and Diarization",

abstract = "Previous work has encouraged domain-invariance in deep speaker embedding by adversarially classifying the dataset or labelled environment to which the generated features belong. We propose a training strategy which aims to produce features that are invariant at the granularity of the recording or channel, a finer grained objective than dataset- or environment- invariance. By training an adversary to predict whether pairs of same-speaker embeddings belong to the same recording in a Siamese fashion, learned features are discouraged from utilizing channel information that may be speaker discriminative during training. Experiments for verification on VoxCeleb and diarization and verification on CALLHOME show promising improvements over a strong baseline in addition to outperforming a dataset-adversarial model. The VoxCeleb model in particular performs well, achieving a 4% relative improvement in EER over a Kaldi baseline, while using a similar architecture and less training data.",

keywords = "Speaker verification, diarization, domain adversarial training, adversarial learning, deep neural network",

author = "Chau Luu and Peter Bell and Steve Renals",

year = "2020",

month = may,

day = "14",

doi = "10.1109/ICASSP40776.2020.9053323",

language = "English",

isbn = "978-1-5090-6632-2",

publisher = "Institute of Electrical and Electronics Engineers (IEEE)",

pages = "7094--7098",

booktitle = "ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)",

address = "United States",

note = "2020 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2020 ; Conference date: 04-05-2020 Through 08-05-2020",

}

Luu, C, Bell, P & Renals, S 2020, Channel Adversarial Training for Speaker Verification and Diarization. in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Institute of Electrical and Electronics Engineers (IEEE), pp. 7094-7098, 2020 IEEE International Conference on Acoustics, Speech, and Signal Processing, Barcelona, Spain, 4/05/20. https://doi.org/10.1109/ICASSP40776.2020.9053323

Channel Adversarial Training for Speaker Verification and Diarization. / Luu, Chau; Bell, Peter ; Renals, Steve.

ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Institute of Electrical and Electronics Engineers (IEEE), 2020. p. 7094-7098.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Channel Adversarial Training for Speaker Verification and Diarization

AU - Luu, Chau

AU - Bell, Peter

AU - Renals, Steve

N1 - Conference code: 45

PY - 2020/5/14

Y1 - 2020/5/14

N2 - Previous work has encouraged domain-invariance in deep speaker embedding by adversarially classifying the dataset or labelled environment to which the generated features belong. We propose a training strategy which aims to produce features that are invariant at the granularity of the recording or channel, a finer grained objective than dataset- or environment- invariance. By training an adversary to predict whether pairs of same-speaker embeddings belong to the same recording in a Siamese fashion, learned features are discouraged from utilizing channel information that may be speaker discriminative during training. Experiments for verification on VoxCeleb and diarization and verification on CALLHOME show promising improvements over a strong baseline in addition to outperforming a dataset-adversarial model. The VoxCeleb model in particular performs well, achieving a 4% relative improvement in EER over a Kaldi baseline, while using a similar architecture and less training data.

AB - Previous work has encouraged domain-invariance in deep speaker embedding by adversarially classifying the dataset or labelled environment to which the generated features belong. We propose a training strategy which aims to produce features that are invariant at the granularity of the recording or channel, a finer grained objective than dataset- or environment- invariance. By training an adversary to predict whether pairs of same-speaker embeddings belong to the same recording in a Siamese fashion, learned features are discouraged from utilizing channel information that may be speaker discriminative during training. Experiments for verification on VoxCeleb and diarization and verification on CALLHOME show promising improvements over a strong baseline in addition to outperforming a dataset-adversarial model. The VoxCeleb model in particular performs well, achieving a 4% relative improvement in EER over a Kaldi baseline, while using a similar architecture and less training data.

KW - Speaker verification

KW - diarization

KW - domain adversarial training

KW - adversarial learning

KW - deep neural network

U2 - 10.1109/ICASSP40776.2020.9053323

DO - 10.1109/ICASSP40776.2020.9053323

M3 - Conference contribution

SN - 978-1-5090-6632-2

SP - 7094

EP - 7098

BT - ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

PB - Institute of Electrical and Electronics Engineers (IEEE)

T2 - 2020 IEEE International Conference on Acoustics, Speech, and Signal Processing

Y2 - 4 May 2020 through 8 May 2020

ER -