The Relevance of the Source Language in Transfer Learning for ASR
This study presents new experiments on Zyrian Komi speech recognition. We use Deep-Speech to train ASR models from a language documentation corpus that contains both contemporary and archival recordings. Earlier studies have shown that transfer learning from English and using a domain matching Komi...
Published in: | Proceedings of the Workshop on Computational Methods for Endangered Languages |
---|---|
Main Authors: | , , , |
Format: | Article in Journal/Newspaper |
Language: | English |
Published: |
Proceedings of the Workshop on Computational Methods for Endangered Languages
2021
|
Subjects: | |
Online Access: | https://journals.colorado.edu/index.php/computel/article/view/959 https://doi.org/10.33011/computel.v1i.959 |
id |
ftucoloradobould:oai:journals.colorado.edu:article/959 |
---|---|
record_format |
openpolar |
spelling |
ftucoloradobould:oai:journals.colorado.edu:article/959 2023-05-15T17:05:11+02:00 The Relevance of the Source Language in Transfer Learning for ASR Hjortnæs, Nils Partanen, Niko Rießler, Michael Tyers, Francis M. 2021-03-02 application/pdf https://journals.colorado.edu/index.php/computel/article/view/959 https://doi.org/10.33011/computel.v1i.959 eng eng Proceedings of the Workshop on Computational Methods for Endangered Languages https://journals.colorado.edu/index.php/computel/article/view/959/889 https://journals.colorado.edu/index.php/computel/article/view/959 doi:10.33011/computel.v1i.959 Proceedings of the Workshop on Computational Methods for Endangered Languages; Vol. 1 (2021): Proceedings of the 4th Workshop on Computational Methods for Endangered Languages; 63-69 10.33011/computel.v1i info:eu-repo/semantics/article info:eu-repo/semantics/publishedVersion Paper 2021 ftucoloradobould https://doi.org/10.33011/computel.v1i.959 https://doi.org/10.33011/computel.v1i 2022-10-18T09:18:49Z This study presents new experiments on Zyrian Komi speech recognition. We use Deep-Speech to train ASR models from a language documentation corpus that contains both contemporary and archival recordings. Earlier studies have shown that transfer learning from English and using a domain matching Komi language model both improve the CER and WER. In this study we experiment with transfer learning from a more relevant source language, Russian, and including Russian text in the language model construction. The motivation for this is that Russian and Komi are contemporary contact languages, and Russian is regularly present in the corpus. We found that despite the close contact of Russian and Komi, the size of the English speech corpus yielded greater performance when used as the source language. Additionally, we can report that already an update in DeepSpeech version improved the CER by 3.9% against the earlier studies, which is an important step in the development of Komi ASR. Article in Journal/Newspaper Komi language University of Colorado Boulder Open Journals Proceedings of the Workshop on Computational Methods for Endangered Languages 1 2 |
institution |
Open Polar |
collection |
University of Colorado Boulder Open Journals |
op_collection_id |
ftucoloradobould |
language |
English |
description |
This study presents new experiments on Zyrian Komi speech recognition. We use Deep-Speech to train ASR models from a language documentation corpus that contains both contemporary and archival recordings. Earlier studies have shown that transfer learning from English and using a domain matching Komi language model both improve the CER and WER. In this study we experiment with transfer learning from a more relevant source language, Russian, and including Russian text in the language model construction. The motivation for this is that Russian and Komi are contemporary contact languages, and Russian is regularly present in the corpus. We found that despite the close contact of Russian and Komi, the size of the English speech corpus yielded greater performance when used as the source language. Additionally, we can report that already an update in DeepSpeech version improved the CER by 3.9% against the earlier studies, which is an important step in the development of Komi ASR. |
format |
Article in Journal/Newspaper |
author |
Hjortnæs, Nils Partanen, Niko Rießler, Michael Tyers, Francis M. |
spellingShingle |
Hjortnæs, Nils Partanen, Niko Rießler, Michael Tyers, Francis M. The Relevance of the Source Language in Transfer Learning for ASR |
author_facet |
Hjortnæs, Nils Partanen, Niko Rießler, Michael Tyers, Francis M. |
author_sort |
Hjortnæs, Nils |
title |
The Relevance of the Source Language in Transfer Learning for ASR |
title_short |
The Relevance of the Source Language in Transfer Learning for ASR |
title_full |
The Relevance of the Source Language in Transfer Learning for ASR |
title_fullStr |
The Relevance of the Source Language in Transfer Learning for ASR |
title_full_unstemmed |
The Relevance of the Source Language in Transfer Learning for ASR |
title_sort |
relevance of the source language in transfer learning for asr |
publisher |
Proceedings of the Workshop on Computational Methods for Endangered Languages |
publishDate |
2021 |
url |
https://journals.colorado.edu/index.php/computel/article/view/959 https://doi.org/10.33011/computel.v1i.959 |
genre |
Komi language |
genre_facet |
Komi language |
op_source |
Proceedings of the Workshop on Computational Methods for Endangered Languages; Vol. 1 (2021): Proceedings of the 4th Workshop on Computational Methods for Endangered Languages; 63-69 10.33011/computel.v1i |
op_relation |
https://journals.colorado.edu/index.php/computel/article/view/959/889 https://journals.colorado.edu/index.php/computel/article/view/959 doi:10.33011/computel.v1i.959 |
op_doi |
https://doi.org/10.33011/computel.v1i.959 https://doi.org/10.33011/computel.v1i |
container_title |
Proceedings of the Workshop on Computational Methods for Endangered Languages |
container_volume |
1 |
container_issue |
2 |
_version_ |
1766059592862662656 |