Multilingual Speech Translation KIT @ IWSLT2021

Proceedings of the 18th International Conference on Spoken Language Translation, pages 154–159 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 154 Multilingual Speech Translation KIT @ IWSLT2021 Ngoc-Quan Pham, Dan He, Tuan-Nam Nguyen, Thanh-Le Ha, Sebastian St¨uker, Alexander Waibel Karlsruhe Institute of Technology ngoc.pham@kit.edu Abstract This paper contains the description for the sub- mission of Karlsruhe Institute of Technology (KIT) for the multilingual TEDx translation task in the IWSLT 2021 evaluation campaign. Our main approach is to develop both cascade and end-to-end systems and eventually com- bine them together to achieve the best possible results for this extremely low-resource setting. The report also confirms certain consistent ar- chitectural improvement added to the Trans- former architecture, for all tasks: translation, transcription and speech translation. 1 Introduction The neural sequence-to-sequence models have rev- olutionalised both automatic speech recognition (ASR) and machine translation in many different aspects, from performance (Luong et al., 2015; Pham et al., 2019a) to various forms such as multi- modal (Barrault et al., 2018) and multilingual (Kan- nan et al., 2019; Ha et al., 2016; Johnson et al., 2016). After multilingual text translation has been established, the recent focus is naturally shifted to multilingual speech translation especially with a series of public speech corpora with multiple trans- lation being released (Iranzo-S´anchez et al., 2020; Wang et al., 2020; Salesky et al., 2021). Recent evaluation campaigns in speech transla- tion have seen a fierce competition between tra- ditional cascade systems and end-to-end counter- parts (Jan et al., 2018, 2019; Ansari et al., 2020). The competition without a doubt would continue in multilingual speech translation especially in a low-resource condition. However, the competition between two modeling schemes suggests that each of them possesses its own strengths and advantages. Notably the cascade models can easily benefit from the separated optimized architectures of each sub- task and enjoy the larger available datasets, while the end-to-end models can theoretically avoid error propagation. This manuscript describes the translation system for the multilingual TEDx task with the aim of combining the strong points of both approaches. We showed that optimizing the cascade models is necessary to bootstrap a powerful end-to-end model, while in the end combining their powers based on ensembling gives promising results. 2 Dataset overview The Multilingual TEDx corpus (Salesky et al., 2021) provided us with the 5 languages Spanish (es), French (fr), Italian (it), Portuguese (pt) and English (en). While speech audio is available for the first 4 languages, text translation is available for all 20 language pairs, and the speech translation parallel data is largely more scarce than the other two. The data statistics is shown in Table 1 and 2. Source → Target en es fr it pt es 36K 102K 3.6K 5.6K 21K fr 30K 20K 116K - - it - - - 50K - pt - 30K - - 90K Table 1: Data statistics for speech recogni- tion/translation in the number of utterances. Source → Target en es fr it pt en - 36.2K 30.5K - …

Content truncated to 500 words