Speaker adaptive training using model agnostic meta-learning

Ondrej Klejch; Joachim Fainberg; Peter Bell; Steve Renals

doi:10.1109/ASRU46091.2019.9003751

Speaker adaptive training using model agnostic meta-learning

Ondrej Klejch, Joachim Fainberg, Peter Bell, Steve Renals

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

Abstract

Speaker adaptive training (SAT) of neural network acoustic models learns models in a way that makes them more suitable for adaptation to test conditions. Conventionally, model-based speaker adaptive training is performed by having a set of speaker dependent parameters that are jointly optimised with speaker independent parameters in order to remove speaker variation. However, this does not scale well if all neural network weights are to be adapted to the speaker. In this paper we formulate speaker adaptive training as a meta-learning task, in which an adaptation process using gradient descent is encoded directly into the training of the model. We compare our approach with test-only adaptation of a standard baseline model and a SAT-LHUC model with a learned speaker adaptation schedule and demonstrate that the meta-learning approach achieves comparable results.

Original language	English
Title of host publication	2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)
Publisher	Institute of Electrical and Electronics Engineers
Pages	881-888
Number of pages	8
ISBN (Electronic)	978-1-7281-0306-8
ISBN (Print)	978-1-7281-0307-5
DOIs	https://doi.org/10.1109/ASRU46091.2019.9003751
Publication status	Published - 20 Feb 2020
Event	IEEE Automatic Speech Recognition and Understanding Workshop 2019 - Sentosa, Singapore Duration: 14 Dec 2019 → 18 Dec 2019 http://asru2019.org/wp/

Conference

Conference	IEEE Automatic Speech Recognition and Understanding Workshop 2019
Abbreviated title	ASRU 2019
Country/Territory	Singapore
City	Sentosa
Period	14/12/19 → 18/12/19
Internet address	http://asru2019.org/wp/

Keywords / Materials (for Non-textual outputs)

speaker adaptation
speaker adaptive training
model-agnostic meta-learning

Access to Document

10.1109/ASRU46091.2019.9003751Licence: All Rights Reserved

Speaker Adaptive Training_KLEJCH_DOA13092019_AFVAccepted author manuscript, 192 KB

https://ieeexplore.ieee.org/document/9003751Licence: All Rights Reserved

ELG- European Language Grid
Germann, U. (Principal Investigator) & Renals, S. (Co-investigator)
EU government bodies
1/01/19 → 30/06/22
Project: Research
SUMMA - Scalable Understanding of Mulitingual Media
Renals, S. (Principal Investigator), Birch-Mayne, A. (Co-investigator) & Cohen, S. (Co-investigator)
EU government bodies
1/02/16 → 31/01/19
Project: Research
Multi-domain speech recognition
Renals, S. (Principal Investigator)
Non-EU industry, commerce and public corporations
1/09/15 → 28/02/19
Project: Research

Cite this

@inproceedings{6cb9b81bd3194fee86e93dad4188aaab,

title = "Speaker adaptive training using model agnostic meta-learning",

abstract = "Speaker adaptive training (SAT) of neural network acoustic models learns models in a way that makes them more suitable for adaptation to test conditions. Conventionally, model-based speaker adaptive training is performed by having a set of speaker dependent parameters that are jointly optimised with speaker independent parameters in order to remove speaker variation. However, this does not scale well if all neural network weights are to be adapted to the speaker. In this paper we formulate speaker adaptive training as a meta-learning task, in which an adaptation process using gradient descent is encoded directly into the training of the model. We compare our approach with test-only adaptation of a standard baseline model and a SAT-LHUC model with a learned speaker adaptation schedule and demonstrate that the meta-learning approach achieves comparable results.",

keywords = "speaker adaptation, speaker adaptive training, model-agnostic meta-learning",

author = "Ondrej Klejch and Joachim Fainberg and Peter Bell and Steve Renals",

year = "2020",

month = feb,

day = "20",

doi = "10.1109/ASRU46091.2019.9003751",

language = "English",

isbn = "978-1-7281-0307-5",

pages = "881--888",

booktitle = "2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)",

publisher = "Institute of Electrical and Electronics Engineers",

address = "United States",

note = "IEEE Automatic Speech Recognition and Understanding Workshop 2019, ASRU 2019 ; Conference date: 14-12-2019 Through 18-12-2019",

url = "http://asru2019.org/wp/",

}

Klejch, O, Fainberg, J, Bell, P & Renals, S 2020, Speaker adaptive training using model agnostic meta-learning. in 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). Institute of Electrical and Electronics Engineers, pp. 881-888, IEEE Automatic Speech Recognition and Understanding Workshop 2019, Sentosa, Singapore, 14/12/19. https://doi.org/10.1109/ASRU46091.2019.9003751

TY - GEN

T1 - Speaker adaptive training using model agnostic meta-learning

AU - Klejch, Ondrej

AU - Fainberg, Joachim

AU - Bell, Peter

AU - Renals, Steve

PY - 2020/2/20

Y1 - 2020/2/20

N2 - Speaker adaptive training (SAT) of neural network acoustic models learns models in a way that makes them more suitable for adaptation to test conditions. Conventionally, model-based speaker adaptive training is performed by having a set of speaker dependent parameters that are jointly optimised with speaker independent parameters in order to remove speaker variation. However, this does not scale well if all neural network weights are to be adapted to the speaker. In this paper we formulate speaker adaptive training as a meta-learning task, in which an adaptation process using gradient descent is encoded directly into the training of the model. We compare our approach with test-only adaptation of a standard baseline model and a SAT-LHUC model with a learned speaker adaptation schedule and demonstrate that the meta-learning approach achieves comparable results.

AB - Speaker adaptive training (SAT) of neural network acoustic models learns models in a way that makes them more suitable for adaptation to test conditions. Conventionally, model-based speaker adaptive training is performed by having a set of speaker dependent parameters that are jointly optimised with speaker independent parameters in order to remove speaker variation. However, this does not scale well if all neural network weights are to be adapted to the speaker. In this paper we formulate speaker adaptive training as a meta-learning task, in which an adaptation process using gradient descent is encoded directly into the training of the model. We compare our approach with test-only adaptation of a standard baseline model and a SAT-LHUC model with a learned speaker adaptation schedule and demonstrate that the meta-learning approach achieves comparable results.

KW - speaker adaptation

KW - speaker adaptive training

KW - model-agnostic meta-learning

U2 - 10.1109/ASRU46091.2019.9003751

DO - 10.1109/ASRU46091.2019.9003751

M3 - Conference contribution

SN - 978-1-7281-0307-5

SP - 881

EP - 888

BT - 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)

PB - Institute of Electrical and Electronics Engineers

T2 - IEEE Automatic Speech Recognition and Understanding Workshop 2019

Y2 - 14 December 2019 through 18 December 2019

ER -

Speaker adaptive training using model agnostic meta-learning

Abstract

Conference

Keywords / Materials (for Non-textual outputs)

Access to Document

Fingerprint

Projects

ELG- European Language Grid

SUMMA - Scalable Understanding of Mulitingual Media

Multi-domain speech recognition

Cite this