WEKO3
アイテム
Winning Tickets in Large Pretrained Speech Models: Exploring the Lottery Ticket Hypothesis in XLS-R
https://ipsj.ixsq.nii.ac.jp/records/2000477
https://ipsj.ixsq.nii.ac.jp/records/2000477492c9763-a396-426f-82bc-742512313599
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
IPSJ-SLP25155004.pdf (788.0 KB)
2027年2月23日からダウンロード可能です。
|
Copyright (c) 2025 by the Information Processing Society of Japan
|
|
| 非会員：¥660, IPSJ:学会員：¥330, SLP:会員：¥0, DLIB:会員：¥0 | ||
| Item type | SIG Technical Reports(1) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 公開日 | 2025-02-23 | |||||||||
| タイトル | ||||||||||
| 言語 | ja | |||||||||
| タイトル | Winning Tickets in Large Pretrained Speech Models: Exploring the Lottery Ticket Hypothesis in XLS-R | |||||||||
| タイトル | ||||||||||
| 言語 | en | |||||||||
| タイトル | Winning Tickets in Large Pretrained Speech Models: Exploring the Lottery Ticket Hypothesis in XLS-R | |||||||||
| 言語 | ||||||||||
| 言語 | eng | |||||||||
| 資源タイプ | ||||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_18gh | |||||||||
| 資源タイプ | technical report | |||||||||
| 著者所属 | ||||||||||
| Graduate School of Informatics, Kyoto University | ||||||||||
| 著者所属 | ||||||||||
| Graduate School of Informatics, Kyoto University | ||||||||||
| 著者所属(英) | ||||||||||
| en | ||||||||||
| Graduate School of Informatics, Kyoto University | ||||||||||
| 著者所属(英) | ||||||||||
| en | ||||||||||
| Graduate School of Informatics, Kyoto University | ||||||||||
| 著者名 |
Jaeyoung,Lee
× Jaeyoung,Lee
× Tatsuya,Kawahara
|
|||||||||
| 著者名(英) |
Jaeyoung Lee
× Jaeyoung Lee
× Tatsuya Kawahara
|
|||||||||
| 論文抄録 | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | Large pretrained ASR models achieve state-of-the-art performance but are computationally expensive. The Lottery Ticket Hypothesis (LTH) hypothesizes that there exist sparse subnetworks, or “winning tickets,” that can match the performance of the full model. This study applies LTH to a large pretrained ASR model, namely XLS-R, demonstrating that winning tickets exist at up to 60% sparsity while maintaining accuracy. Using a subset of Common Voice covering 90 languages, we find that moderate pruning (80%-51% of the model size) enhances generalization, consistent with prior LTH findings. Our results confirm LTH's applicability to large ASR models, opening new avenues for efficient and scalable speech recognition. | |||||||||
| 論文抄録(英) | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | Large pretrained ASR models achieve state-of-the-art performance but are computationally expensive. The Lottery Ticket Hypothesis (LTH) hypothesizes that there exist sparse subnetworks, or “winning tickets,” that can match the performance of the full model. This study applies LTH to a large pretrained ASR model, namely XLS-R, demonstrating that winning tickets exist at up to 60% sparsity while maintaining accuracy. Using a subset of Common Voice covering 90 languages, we find that moderate pruning (80%-51% of the model size) enhances generalization, consistent with prior LTH findings. Our results confirm LTH's applicability to large ASR models, opening new avenues for efficient and scalable speech recognition. | |||||||||
| 書誌レコードID | ||||||||||
| 収録物識別子タイプ | NCID | |||||||||
| 収録物識別子 | AN10442647 | |||||||||
| 書誌情報 |
研究報告音声言語情報処理（SLP） 巻 2025-SLP-155, 号 4, p. 1-4, 発行日 2025-02-23 |
|||||||||
| ISSN | ||||||||||
| 収録物識別子タイプ | ISSN | |||||||||
| 収録物識別子 | 2188-8663 | |||||||||
| Notice | ||||||||||
| SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc. | ||||||||||
| 出版者 | ||||||||||
| 言語 | ja | |||||||||
| 出版者 | 情報処理学会 | |||||||||
