BUT Systems for IWSLT 2023 Marathi - Hindi Low Resource Speech Translation Task
20558Authors
nan
Abstract
AbstractThis paper describes the systems submitted for Marathi to Hindi low-resource speech translation task. Our primary submission is based on an end-to-end direct speech translation system, whereas the contrastive one is a cascaded system. The backbone of both the systems is a Hindi-Marathi bilingual ASR system trained on 2790 hours of imperfect transcribed speech. The end-to-end speech translation system was directly initialized from the ASR, and then fine-tuned for direct speech translation with an auxiliary CTC loss for translation. The MT model for the cascaded system is initialized from a cross-lingual language model, which was then fine-tuned using 1.6 M parallel sentences. All our systems were trained from scratch on publicly available datasets. In the end, we use a language model to re-score the n-best hypotheses. Our primary submission achieved 30.5 and 39.6 BLEU whereas the contrastive system obtained 21.7 and 28.6 BLEU on official dev and test sets respectively. The paper also presents the analysis on several experiments that were conducted and outlines the strategies for improving speech translation in low-resource scenarios.
Keywords
Full Content
Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 1–61 July 13-14, 2023 ©2023 Association for Computational Linguistics FINDINGS OF THE IWSLT 2023 EVALUATION CAMPAIGN Milind Agarwal1 Sweta Agrawal2 Antonios Anastasopoulos1 Luisa Bentivogli3 Ondˇrej Bojar4 Claudia Borg5 Marine Carpuat2 Roldano Cattoni3 Mauro Cettolo3 Mingda Chen6 William Chen7 Khalid Choukri8 Alexandra Chronopoulou9 Anna Currey10 Thierry Declerck11 Qianqian Dong12 Kevin Duh13 Yannick Est`eve14 Marcello Federico10 Souhir Gahbiche15 Barry Haddow16 Benjamin Hsu10 Phu Mon Htut10 Hirofumi Inaguma6 D´avid Javorsk´y4 John Judge17 Yasumasa Kano18 Tom Ko12 Rishu Kumar4 Pengwei Li6 Xutai Ma6 Prashant Mathur10 Evgeny Matusov19 Paul McNamee13 John P. McCrae20 Kenton Murray13 Maria Nadejde10 Satoshi Nakamura18 Matteo Negri3 Ha Nguyen14 Jan Niehues21 Xing Niu10 Atul Kr. Ojha20 John E. Ortega22 Proyag Pal16 Juan Pino6 Lonneke van der Plas23 Peter Pol´ak4 Elijah Rippeth2 Elizabeth Salesky13 Jiatong Shi7 Matthias Sperber24 Sebastian St¨uker25 Katsuhito Sudoh18 Yun Tang6 Brian Thompson10 Kevin Tran6 Marco Turchi25 Alex Waibel7 Mingxuan Wang12 Shinji Watanabe7 Rodolfo Zevallos26 1GMU 2UMD 3FBK 4Charles U. 5U. Malta 6Meta 7CMU 8ELDA 9LMU 10AWS 11DFKI 12ByteDance 13JHU 14Avignon U. 15Airbus 16U. Edinburgh 18NAIST 19AppTek 20U. Galway 21KIT 22Northeastern U. 23IDIAP 24Apple 25Zoom 26U. Pompeu Fabra Abstract This paper reports on the shared tasks orga- nized by the 20th IWSLT Conference. The shared tasks address 9 scientific challenges in spoken language translation: simultane- ous and offline translation, automatic subti- tling and dubbing, speech-to-speech transla- tion, multilingual, dialect and low-resource speech translation, and formality control. The shared tasks attracted a total of 38 submis- sions by 31 teams. The growing interest to- wards spoken language translation is also wit- nessed by the constantly increasing number of shared task organizers and contributors to the overview paper, almost evenly distributed across industry and academia. 1 Introduction The International Conference on Spoken Lan- guage Translation (IWSLT) is the premier an- nual scientific conference for all aspects of spoken language translation (SLT). IWSLT is organized by the Special Interest Group on Spoken Lan- guage Translation (SIG-SLT), which is supported by ACL, ISCA and ELRA. Like in all previous editions (Akiba et al., 2004; Eck and Hori, 2005; Paul, 2006; Fordyce, 2007; Paul, 2008, 2009; Paul et al., 2010; Federico et al., 2011, 2012; Cettolo et al., 2013, 2014, 2015, 2016, 2017; Niehues et al., 2018, 2019; Ansari et al., 2020; Anasta- sopoulos et al., 2021, 2022),this year’s conference was preceded by an evaluation campaign featur- ing shared tasks addressing scientific challenges in SLT. This paper reports on the 2023 IWSLT Eval- uation Campaign, which offered the following 9 shared tasks: • Offline SLT, with focus on speech-to-text translation of recorded conferences and inter- views from English to German, Japanese and Chinese. • Simultaneous SLT, focusing on speech-to- text translation of streamed audio of confer- ences and interviews from English to German, Japanese and Chinese. • Automatic Subtitling, with focus on speech- to-subtitle translation of audio-visual docu- ments from English to German and Spanish. • Multilingual SLT, with focus on speech-to- text translation of recorded scientific talks from 1 Team …
Content truncated to 500 words
Labeled Text
AbstractThis paper describes the systems submitted for Marathi to Hindi low-resource speech translation task. Our primary submission is based on an end-to-end direct speech translation system, whereas the contrastive one is a cascaded system. The backbone of both the systems is a Hindi-Marathi bilingual ASR system trained on 2790 hours of imperfect transcribed speech. The end-to-end speech translation system was directly initialized from the ASR, and then fine-tuned for direct speech translation with an auxiliary CTC loss for translation. The MT model for the cascaded system is initialized from a cross-lingual language model, which was then fine-tuned using 1.6 M parallel sentences. All our systems were trained from scratch on publicly available datasets. In the end, we use a language model to re-score the n-best hypotheses. Our primary submission achieved 30.5 and 39.6 BLEU whereas the contrastive system obtained 21.7 and 28.6 BLEU on official dev and test sets respectively. The paper also presents the analysis on several experiments that were conducted and outlines the strategies for improving speech translation in low-resource scenarios.
Publication Details
- Paper ID:
- 20558
- Year:
- 2023
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan