MuST-C: a Multilingual Speech Translation Corpus

Proceedings of NAACL-HLT 2019, pages 2012–2017 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2012 MuST-C: a Multilingual Speech Translation Corpus Mattia Antonino Di Gangi1,2, Roldano Cattoni1, Luisa Bentivogli1, Matteo Negri1 and Marco Turchi1 1Fondazione Bruno Kessler 2University of Trento Trento, Italy {digangi,cattoni,bentivo,negri,turchi}@fbk.eu Abstract Current research on spoken language transla- tion (SLT) has to confront with the scarcity of sizeable and publicly available training cor- pora. This problem hinders the adoption of neural end-to-end approaches, which represent the state of the art in the two parent tasks of SLT: automatic speech recognition and ma- chine translation. To fill this gap, we cre- ated MuST-C, a multilingual speech transla- tion corpus whose size and quality will facili- tate the training of end-to-end systems for SLT from English into 8 languages. For each tar- get language, MuST-C comprises at least 385 hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual transcriptions and translations. Together with a description of the corpus creation methodology (scalable to add new data and cover new languages), we provide an empirical verification of its quality and SLT results computed with strong baseline system on each language direction. 1 Introduction Besides the increased computing power, the recent surge of neural end-to-end approaches to natural language processing tasks has been stoked by the increased availability of data. For instance, when supported by sizeable training corpora, the robust- ness and the strong generalization capabilities of neural networks led to their dominance over previ- ous paradigms both in automatic speech recogni- tion (ASR (Chiu et al., 2018)) and machine trans- lation (MT (Bojar et al., 2018)). Compared to its two parent research areas, spo- ken language translation (SLT) has not shown such a steady progress yet. Despite recent claims by big industry players about the effectiveness of end-to- end learning (Weiss et al., 2017; Jia et al., 2018), its adoption does not yet represent the mainstream solution to the SLT task. One of the main obstacles Corpus Languages Hours Niehues et al. (2018) En→De 273 Kocabiyikoglu et al. (2018) En→Fr 236 Tohyama et al. (2005) En↔Jp 182 Paulik and Waibel (2009) En→Es 111 Es→En 105 Post et al. (2013) En→Es 38 St¨uker et al. (2012) De→En 37 Shimizu et al. (2014) En↔Jp 22 Federmann and Lewis (2017) En↔Jp/Zh 22 Bendazzoli and Sandrelli (2005) En↔It/Es 18 It↔Es B´erard et al. (2016) Fr→En 17 Federmann and Lewis (2016) En↔Fr/De 8 Woldeyohannis et al. (2017) Am→En 7 Godard et al. (2017) Mboshi→Fr 4 Table 1: Publicly available SLT corpora. The two most recent resources (also known as IWSLT18 and Aug- mented LibriSpeech) are also the largest ones. Though considerably smaller, the Fisher and Callhome cor- pus described in (Post et al., 2013) is among the most widely used ones in previous research. to a stable dominance of the end-to-end paradigm also in this area is the scarcity of training corpora. While cascade ASR+MT solutions can exploit the wealth of task-specific …

Content truncated to 500 words