A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic

Dan Wells; Korin Richmond; Will Lamb

doi:10.21437/Interspeech.2023-2056

A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic

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

Abstract / Description of output

In this work we present an end-to-end pipeline for building a speech corpus and text-to-speech synthesis system for a new language without reference to any expert-defined linguistic resources. We segment and align over 85 hours of Scottish Gaelic recordings found online and select 2- and 8-hour subsets with comprehensive coverage of speech sounds based on self-supervised discrete acoustic unit sequences. We then compare FastPitch models trained on these relatively small data sets using character, acoustic unit and phone inputs. According to native speaker listening test judgements, characters serve well for Gaelic given its regular orthography, even in these limited data scenarios. We release our corpus building recipe so that others may easily apply our work to new languages.

Original language	English
Title of host publication	Proceedings of the Annual Conference of the International Speech Communication Association
Subtitle of host publication	Interspeech 2023
Editors	Naomi Harte, Julie Carson-Berndsen, Gareth Jones
Place of Publication	Dublin
Publisher	ISCA
Pages	4324-4328
DOIs	https://doi.org/10.21437/Interspeech.2023-2056
Publication status	Published - Sept 2023
Event	Interspeech 2023 - Dublin, Ireland Duration: 20 Aug 2023 → 24 Aug 2023 Conference number: 24 https://www.interspeech2023.org/

Publication series

Name	Interspeech - Annual Conference of the International Speech Communication Association
Publisher	ISCA
ISSN (Electronic)	2308-457X

Conference

Conference	Interspeech 2023
Country/Territory	Ireland
City	Dublin
Period	20/08/23 → 24/08/23
Internet address	https://www.interspeech2023.org/

Keywords / Materials (for Non-textual outputs)

Scottish Gaelic
speech synthesis
low-resource
speech corpus creation
found data

Access to Document

10.21437/Interspeech.2023-2056Licence: Creative Commons: Attribution (CC-BY)

WellsEtAl2023ALow-ResourcePipelineFinal published version, 313 KBLicence: Creative Commons: Attribution (CC-BY)

Cite this

Wells, D., Richmond, K., & Lamb, W. (2023). A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic. In N. Harte, J. Carson-Berndsen, & G. Jones (Eds.), Proceedings of the Annual Conference of the International Speech Communication Association: Interspeech 2023 (pp. 4324-4328). (Interspeech - Annual Conference of the International Speech Communication Association). ISCA. https://doi.org/10.21437/Interspeech.2023-2056

Wells, Dan ; Richmond, Korin ; Lamb, Will. / A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic. Proceedings of the Annual Conference of the International Speech Communication Association: Interspeech 2023. editor / Naomi Harte ; Julie Carson-Berndsen ; Gareth Jones. Dublin : ISCA, 2023. pp. 4324-4328 (Interspeech - Annual Conference of the International Speech Communication Association).

@inproceedings{048ae96369a742d78ba6533ed7febbe1,

title = "A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic",

abstract = "In this work we present an end-to-end pipeline for building a speech corpus and text-to-speech synthesis system for a new language without reference to any expert-defined linguistic resources. We segment and align over 85 hours of Scottish Gaelic recordings found online and select 2- and 8-hour subsets with comprehensive coverage of speech sounds based on self-supervised discrete acoustic unit sequences. We then compare FastPitch models trained on these relatively small data sets using character, acoustic unit and phone inputs. According to native speaker listening test judgements, characters serve well for Gaelic given its regular orthography, even in these limited data scenarios. We release our corpus building recipe so that others may easily apply our work to new languages.",

keywords = "Scottish Gaelic, speech synthesis, low-resource, speech corpus creation, found data",

author = "Dan Wells and Korin Richmond and Will Lamb",

note = "Acknowledgements: We would like to thank Ruairidh MacIlleathain for allowing us to use his voice for this work. We also thank BBC ALBA and MG ALBA for granting permission as copyright holder and funder of the Letter to Learners series of recordings, respectively. This work was supported in part by the UKRI Centre for Doctoral Training in Natural Language Processing, funded by the UKRI (grant EP/S022481/1) and the University of Edinburgh, School of Informatics and School of Philosophy, Psychology & Language Sciences.; Interspeech 2023 ; Conference date: 20-08-2023 Through 24-08-2023",

year = "2023",

month = sep,

doi = "10.21437/Interspeech.2023-2056",

language = "English",

series = "Interspeech - Annual Conference of the International Speech Communication Association",

publisher = "ISCA",

pages = "4324--4328",

editor = "Naomi Harte and Julie Carson-Berndsen and Gareth Jones",

booktitle = "Proceedings of the Annual Conference of the International Speech Communication Association",

url = "https://www.interspeech2023.org/",

}

Wells, D, Richmond, K & Lamb, W 2023, A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic. in N Harte, J Carson-Berndsen & G Jones (eds), Proceedings of the Annual Conference of the International Speech Communication Association: Interspeech 2023. Interspeech - Annual Conference of the International Speech Communication Association, ISCA, Dublin, pp. 4324-4328, Interspeech 2023, Dublin, Ireland, 20/08/23. https://doi.org/10.21437/Interspeech.2023-2056

A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic. / Wells, Dan; Richmond, Korin ; Lamb, Will.
Proceedings of the Annual Conference of the International Speech Communication Association: Interspeech 2023. ed. / Naomi Harte; Julie Carson-Berndsen; Gareth Jones. Dublin: ISCA, 2023. p. 4324-4328 (Interspeech - Annual Conference of the International Speech Communication Association).

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

TY - GEN

T1 - A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic

AU - Wells, Dan

AU - Richmond, Korin

AU - Lamb, Will

N1 - Conference code: 24

PY - 2023/9

Y1 - 2023/9

N2 - In this work we present an end-to-end pipeline for building a speech corpus and text-to-speech synthesis system for a new language without reference to any expert-defined linguistic resources. We segment and align over 85 hours of Scottish Gaelic recordings found online and select 2- and 8-hour subsets with comprehensive coverage of speech sounds based on self-supervised discrete acoustic unit sequences. We then compare FastPitch models trained on these relatively small data sets using character, acoustic unit and phone inputs. According to native speaker listening test judgements, characters serve well for Gaelic given its regular orthography, even in these limited data scenarios. We release our corpus building recipe so that others may easily apply our work to new languages.

AB - In this work we present an end-to-end pipeline for building a speech corpus and text-to-speech synthesis system for a new language without reference to any expert-defined linguistic resources. We segment and align over 85 hours of Scottish Gaelic recordings found online and select 2- and 8-hour subsets with comprehensive coverage of speech sounds based on self-supervised discrete acoustic unit sequences. We then compare FastPitch models trained on these relatively small data sets using character, acoustic unit and phone inputs. According to native speaker listening test judgements, characters serve well for Gaelic given its regular orthography, even in these limited data scenarios. We release our corpus building recipe so that others may easily apply our work to new languages.

KW - Scottish Gaelic

KW - speech synthesis

KW - low-resource

KW - speech corpus creation

KW - found data

UR - https://github.com/dan-wells/kiss-aligner/tree/main/egs/learngaelic_litir

UR - https://learngaelic.scot/litir/

U2 - 10.21437/Interspeech.2023-2056

DO - 10.21437/Interspeech.2023-2056

M3 - Conference contribution

T3 - Interspeech - Annual Conference of the International Speech Communication Association

SP - 4324

EP - 4328

BT - Proceedings of the Annual Conference of the International Speech Communication Association

A2 - Harte, Naomi

A2 - Carson-Berndsen, Julie

A2 - Jones, Gareth

PB - ISCA

CY - Dublin

T2 - Interspeech 2023

Y2 - 20 August 2023 through 24 August 2023

ER -

Wells D, Richmond K , Lamb W. A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic. In Harte N, Carson-Berndsen J, Jones G, editors, Proceedings of the Annual Conference of the International Speech Communication Association: Interspeech 2023. Dublin: ISCA. 2023. p. 4324-4328. (Interspeech - Annual Conference of the International Speech Communication Association). Epub 2023 Aug 24. doi: 10.21437/Interspeech.2023-2056

A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic

Abstract / Description of output

Publication series

Conference

Keywords / Materials (for Non-textual outputs)

Access to Document

Other files and links

Fingerprint

Two studentships in Natural Language Processing

Cite this

A low-resource pipeline for text-to-speech from found data with application to Scottish Gaelic

Abstract / Description of output

Publication series

Conference

Keywords / Materials (for Non-textual outputs)

Access to Document

Other files and links

Fingerprint

Projects

Two studentships in Natural Language Processing

Cite this