Elamur Research
☰
home / papers / 2019 / MuST-C: a Multilingual Speech Translation Corpus

MuST-C: a Multilingual Speech Translation Corpus

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/N19-1202.pdf

Summary

End-to-end neural spoken language translation (SLT) has struggled to match its success in ASR and MT, mainly because public speech-to-translation corpora are scarce, cover few languages, and are often under 100 hours. The authors address this with MuST-C, a corpus for translating English speech into 8 languages, built from English TED Talks and their volunteer-made transcripts and translations.

Their pipeline first aligns transcripts and translations at the sentence level using Gargantua. It then maps the English sentences onto the audio with the Gentle forced aligner, which is based on Kaldi. Noisy talks and segments are filtered out according to how many words failed to align.

To check quality, they trained models on the IWSLT18 En-De corpus and on MuST-C data drawn from the same talks. The MuST-C versions scored better on ASR, MT and SLT. They also report baselines for every language direction.

MuST-C gives the field a large, reusable benchmark with fixed splits, making end-to-end SLT practical without proprietary data.

Key points

  • Coverage and size: MuST-C pairs English source audio with eight target languages: German, Russian, Romanian, Portuguese, Dutch, Italian, French and Spanish. Each section has 385 to 504 hours of speech and roughly 211K to 280K sentences, so every section is larger than any earlier public SLT corpus.
  • Filtering rules: A whole talk is removed if at least 15% of its words go unaligned by Gentle. Any individual sentence with no aligned words is also dropped. Audio is provided as 40-dimensional log Mel filter-bank features extracted with XNMT.
  • Comparison with IWSLT18 (En-De): Training on MuST-C instead of IWSLT18 improved results on all three tasks:
    • ASR WER fell from 42.15 to 32.05.
    • MT BLEU rose from 24.90 to 25.46.
    • SLT BLEU rose from 8.94 to 12.25.
  • Baselines: SLT used a character-level LSTM encoder-decoder in fairseq, adapted from Bérard et al.; MT used ModernMT, a Transformer system. SLT BLEU ranged from 7.22 for Russian to 22.29 for French, and ASR WER stayed between about 26 and 28. The fixed dev set has 1.4K segments from 11 talks, and the test set has 2.5K segments from 27 talks.
  • License and access: The corpus is released under CC BY-NC-ND 4.0 at mustc.fbk.eu. The creation method is designed to scale to new talks and new target languages.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

232 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 230 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n19-1202

@inproceedings{gangi2019must,
  title = {{MuST-C: a Multilingual Speech Translation Corpus}},
  author = {Gangi, Mattia A. Di and Cattoni, Roldano and Bentivogli, Luisa and Negri, Matteo and Turchi, Marco},
  booktitle = {Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)},
  year = {2019},
  doi = {10.18653/v1/n19-1202},
  url = {https://doi.org/10.18653/v1/n19-1202}
}

References

27 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Antonios Anastasopoulos and David Chiang. 2018. Tied Multitask Learning for Neural Speech Translation. In Proceedings of NAACL-HLT 2018, pages 82–91, New Orleans, Louisiana. find
  2. Sameer Bansal, Herman Kamper, Karen Livescu, Adam Lopez, and Sharon Goldwater. 2018. find
  3. Pre-Christian Federmann and William D Lewis. 2016. Microsoft Speech Language Translation (MSLT) Corpus: The IWSLT 2016 Release for English, French and German. In Proceedings of IWSLT 2016, Seattle, USA. find
  4. Christian Federmann and William D Lewis. 2017. The Microsoft Speech Language Translation (MSLT) Corpus for Chinese and Japanese: Conversational Test data for Machine Translation and Speech Recognition. In Proceedings of the 16th Machine Translation Summit, Nagoya, Japan. find
  5. Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. 2017. Convolutional Sequence to Sequence Learning. In Proceedings of ICML 2017, pages 1243–1252, Sydney, Australia. find our card
  6. Pierre Godard, Gilles Adda, Martine Adda-Decker, Juan Benjumea, Laurent Besacier, et al. 2017. A Very Low Resource Language Speech Corpus for Computational Language Documentation Experiments. arXiv preprint arXiv:1710.03501. find
  7. Alex Graves, Abdel-rahman Mohamed, and Geoffrey E. Hinton. 2013. Speech Recognition with Deep Recurrent Neural Networks. In Proceedings of ICASSP 2018, pages 6645–6649, Vancouver, BC, Canada. find
  8. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural computation, 9(8). find
  9. Ye Jia, Melvin Johnson, Wolfgang Macherey, Ron-J. Weiss, Yuan Cao, Chung-Cheng Chiu, Stella-Laurenzo Ari, and Yonghui Wu. 2018. Leveraging Weakly Supervised Data to Improve End-to-End Speech-to-Text Translation. ArXiv e-prints arXiv:1811.02050. — [arXiv:1811.02050](https://arxiv.org/abs/1811.02050) find
  10. Ali Can Kocabiyikoglu, Laurent Besacier, and Olivier Kraif. 2018. Augmenting Librispeech with French Translations: A Multimodal Corpus for Direct Speech Translation Evaluation. In Proceedings of LREC 2018, Miyazaki, Japan. find
  11. Dan Liu, Junhua Liu, Wu Guo, Shifu Xiong, Zhiqiang Ma, Rui Song, Chongliang Wu, and Quan Liu. 2018. The USTC-NEL Speech Translation system at IWSLT 2018. In Proceedings of IWSLT 2018, Bruges, Belgium. find
  12. Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effective Approaches to Attention-based Neural Machine Translation. In Proceedings of EMNLP 2015, pages 1412–1421, Lisbon, Portugal. find
  13. Graham Neubig, Matthias Sperber, Xinyi Wang, Matthieu Felix, Austin Matthews, et al. 2018. XNMT: The eXtensible Neural Machine Translation Toolkit. In Proceedings of AMTA 2018, pages 185– 192, Boston, MA. find
  14. Jan Niehues, Roldano Cattoni, Sebastian Stüker, Mauro Cettolo, Marco Turchi, and Marcello Federico. 2018. find
  15. The IWSLT 2018 Evaluation Campaign. In Proceedings of IWSLT 2018, Bruges, Belgium. find
  16. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a Method for Automatic Evaluation of Machine Translation. In Proceedings of ACL 2002, Philadelphia, PA, USA. find our card
  17. Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, and Yoshua Bengio. 2014. How to Construct Deep Recurrent Neural Networks. In Proceedings of ICLR 2014, Banff, Canada. find
  18. Matthias Paulik and Alex Waibel. 2009. Automatic Translation from Parallel Speech: Simultaneous Interpretation as MT Training Data. In Proceedings of ASRU 2009, pages 496–501, Merano, Italy. find
  19. Matt Post, Gaurav Kumar, Adam Lopez, Damianos Karakos, Chris Callison-Burch, and Sanjeev Khudanpur. 2013. Improved Speech-to-Text Translation with the Fisher and Callhome Spanish–English Speech Translation Corpus. In Proceedings of IWSLT 2013, Heidelberg, Germany. find
  20. Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukáš Burget, Ondřej Glembek, K. Nagendra Goel, Mirko Hannemann, Petr Motlı́ček, Yanmin Qian, Petr Schwarz, Jan Silovský, Georg Stemmer, and Karel Veselý. 2011. The Kaldi Speech Recognition Toolkit. In Proceedings of ASRU 2011, pages 1–4, Big Isla… find
  21. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Neural Machine Translation of Rare Words with Subword Units. arXiv preprint arXiv:1508.07909. find
  22. Hiroaki Shimizu, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2014. Collection of a Simultaneous Translation Corpus for Comparative Analysis. In Proceedings of LREC 2014, Reykjavik, Iceland. find
  23. Sebastian Stüker, Florian Kraft, Christian Mohr, Teresa Herrmann, Eunah Cho, and Alex Waibel. 2012. The KIT Lecture Corpus for Speech Translation. In Proceedings of LREC-2012, Istanbul, Turkey. find
  24. Hitomi Tohyama, Shigeki Matsubara, Nobuo Kawaguchi, and Yasuyoshi Inagaki. 2005. Construction and utilization of bilingual speech corpus for simultaneous machine interpretation research. In Proceedings of INTERSPEECH, pages 1585–1588, Lisbon, Portugal. find
  25. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Proceedings of NIPS 2017, pages 5998–6008, Long Beach, CA, USA. find
  26. Ron J. Weiss, Jan Chorowski, Navdeep Jaitly, Yonghui Wu, and Zhifeng Chen. 2017. Sequence-to-Sequence Models Can Directly Translate Foreign Speech. In Proceedings of Interspeech 2017, pages 2625–2629, Stockholm, Sweden. find
  27. Michael Melese Woldeyohannis, Laurent Besacier, and Million Meshesha. 2017. A Corpus for Amharic-English Speech Translation: the Case of Tourism Domain. In Proceedings of ICT4DA 2017, pages 129–139, Bahir Dar, Ethiopia. find