Large scale weakly and semi-supervised learning for low-resource video ASR

Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Singh, Kritika, Manohar, Vimal, Xiao, Alex, Edunov, Sergey, Girshick, Ross, Liptchinsky, Vitaliy, Fuegen, Christian, Saraf, Yatharth, Zweig, Geoffrey, Mohamed, Abdelrahman
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Singh, Kritika
Manohar, Vimal
Xiao, Alex
Edunov, Sergey
Girshick, Ross
Liptchinsky, Vitaliy
Fuegen, Christian
Saraf, Yatharth
Zweig, Geoffrey
Mohamed, Abdelrahman
description Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.
doi_str_mv 10.48550/arxiv.2005.07850
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2005_07850</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2005_07850</sourcerecordid><originalsourceid>FETCH-LOGICAL-a670-6cd95b044e21163939ea9834775cb935b9e230dba582555fea341991f7c54b33</originalsourceid><addsrcrecordid>eNotz71OwzAUQGEvDKjwAEz4BRzs2DeOx6riV5GQKHt07dxUFm5S2WpK3x5RmM52pI-xOyUr0wLIB8zfcalqKaGStgV5zd46zDviJWAifiL8SmeO08AL7aMoxwPlJRYaeCLMU5x2fJwzT_NJZCrzMQfiSxxo5uvtxw27GjEVuv3vim2fHj83L6J7f37drDuBjZWiCYMDL42hWqlGO-0IXauNtRC80-Ad1VoOHqGtAWAk1EY5p0YbwHitV-z-73qx9Icc95jP_a-pv5j0D7YdRiU</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Large scale weakly and semi-supervised learning for low-resource video ASR</title><source>arXiv.org</source><creator>Singh, Kritika ; Manohar, Vimal ; Xiao, Alex ; Edunov, Sergey ; Girshick, Ross ; Liptchinsky, Vitaliy ; Fuegen, Christian ; Saraf, Yatharth ; Zweig, Geoffrey ; Mohamed, Abdelrahman</creator><creatorcontrib>Singh, Kritika ; Manohar, Vimal ; Xiao, Alex ; Edunov, Sergey ; Girshick, Ross ; Liptchinsky, Vitaliy ; Fuegen, Christian ; Saraf, Yatharth ; Zweig, Geoffrey ; Mohamed, Abdelrahman</creatorcontrib><description>Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.</description><identifier>DOI: 10.48550/arxiv.2005.07850</identifier><language>eng</language><subject>Computer Science - Computation and Language ; Computer Science - Sound</subject><creationdate>2020-05</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2005.07850$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2005.07850$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Singh, Kritika</creatorcontrib><creatorcontrib>Manohar, Vimal</creatorcontrib><creatorcontrib>Xiao, Alex</creatorcontrib><creatorcontrib>Edunov, Sergey</creatorcontrib><creatorcontrib>Girshick, Ross</creatorcontrib><creatorcontrib>Liptchinsky, Vitaliy</creatorcontrib><creatorcontrib>Fuegen, Christian</creatorcontrib><creatorcontrib>Saraf, Yatharth</creatorcontrib><creatorcontrib>Zweig, Geoffrey</creatorcontrib><creatorcontrib>Mohamed, Abdelrahman</creatorcontrib><title>Large scale weakly and semi-supervised learning for low-resource video ASR</title><description>Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.</description><subject>Computer Science - Computation and Language</subject><subject>Computer Science - Sound</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2020</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotz71OwzAUQGEvDKjwAEz4BRzs2DeOx6riV5GQKHt07dxUFm5S2WpK3x5RmM52pI-xOyUr0wLIB8zfcalqKaGStgV5zd46zDviJWAifiL8SmeO08AL7aMoxwPlJRYaeCLMU5x2fJwzT_NJZCrzMQfiSxxo5uvtxw27GjEVuv3vim2fHj83L6J7f37drDuBjZWiCYMDL42hWqlGO-0IXauNtRC80-Ad1VoOHqGtAWAk1EY5p0YbwHitV-z-73qx9Icc95jP_a-pv5j0D7YdRiU</recordid><startdate>20200515</startdate><enddate>20200515</enddate><creator>Singh, Kritika</creator><creator>Manohar, Vimal</creator><creator>Xiao, Alex</creator><creator>Edunov, Sergey</creator><creator>Girshick, Ross</creator><creator>Liptchinsky, Vitaliy</creator><creator>Fuegen, Christian</creator><creator>Saraf, Yatharth</creator><creator>Zweig, Geoffrey</creator><creator>Mohamed, Abdelrahman</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20200515</creationdate><title>Large scale weakly and semi-supervised learning for low-resource video ASR</title><author>Singh, Kritika ; Manohar, Vimal ; Xiao, Alex ; Edunov, Sergey ; Girshick, Ross ; Liptchinsky, Vitaliy ; Fuegen, Christian ; Saraf, Yatharth ; Zweig, Geoffrey ; Mohamed, Abdelrahman</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a670-6cd95b044e21163939ea9834775cb935b9e230dba582555fea341991f7c54b33</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2020</creationdate><topic>Computer Science - Computation and Language</topic><topic>Computer Science - Sound</topic><toplevel>online_resources</toplevel><creatorcontrib>Singh, Kritika</creatorcontrib><creatorcontrib>Manohar, Vimal</creatorcontrib><creatorcontrib>Xiao, Alex</creatorcontrib><creatorcontrib>Edunov, Sergey</creatorcontrib><creatorcontrib>Girshick, Ross</creatorcontrib><creatorcontrib>Liptchinsky, Vitaliy</creatorcontrib><creatorcontrib>Fuegen, Christian</creatorcontrib><creatorcontrib>Saraf, Yatharth</creatorcontrib><creatorcontrib>Zweig, Geoffrey</creatorcontrib><creatorcontrib>Mohamed, Abdelrahman</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Singh, Kritika</au><au>Manohar, Vimal</au><au>Xiao, Alex</au><au>Edunov, Sergey</au><au>Girshick, Ross</au><au>Liptchinsky, Vitaliy</au><au>Fuegen, Christian</au><au>Saraf, Yatharth</au><au>Zweig, Geoffrey</au><au>Mohamed, Abdelrahman</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Large scale weakly and semi-supervised learning for low-resource video ASR</atitle><date>2020-05-15</date><risdate>2020</risdate><abstract>Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.</abstract><doi>10.48550/arxiv.2005.07850</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2005.07850
ispartof
issn
language eng
recordid cdi_arxiv_primary_2005_07850
source arXiv.org
subjects Computer Science - Computation and Language
Computer Science - Sound
title Large scale weakly and semi-supervised learning for low-resource video ASR
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2024-12-24T19%3A55%3A52IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Large%20scale%20weakly%20and%20semi-supervised%20learning%20for%20low-resource%20video%20ASR&rft.au=Singh,%20Kritika&rft.date=2020-05-15&rft_id=info:doi/10.48550/arxiv.2005.07850&rft_dat=%3Carxiv_GOX%3E2005_07850%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true