Maastricht University’s Multilingual Speech Translation System for IWSLT 2021

Proceedings of the 18th International Conference on Spoken Language Translation, pages 138–143 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 138 Maastricht University’s Multilingual Speech Translation System for IWSLT 2021 Danni Liu, Jan Niehues Department of Data Science and Knowledge Engineering, Maastricht University {danni.liu,jan.niehues}@maastrichtuniversity.nl Abstract This paper describes Maastricht University’s participation in the IWSLT 2021 multilingual speech translation track. The task of this track is to build multilingual speech translation sys- tems in supervised and zero-shot directions. Our primary system is an end-to-end model that performs both speech transcription and translation. We observe that the joint training for the two tasks is complementary especially when the speech translation data is scarce. On the source and target side, we use data augmen- tation and pseudo-labels respectively to im- prove the performance of our systems. We also introduce an ensembling technique that con- sistently improves the quality of transcriptions and translations. The experiments show that the end-to-end system is competitive with its cascaded counterpart especially in zero-shot conditions. 1 Introduction In this paper, we describe our systems for the mul- tilingual speech translation track of IWSLT 2021. Speech translation (B´erard et al., 2016; Weiss et al., 2017) is the task of converting speech utterances to their translation in other languages. While “end- to-end” modeling (Di Gangi et al., 2019; Sperber et al., 2019) of the speech translation pipeline has become the dominant approach, an open challenge remains in terms of data scarcity. As the amount of speech directly paired with translation is lower compared to speech transcription or text-to-text translation, it is especially crucial for models to be data-efficient. In this context, multilingual speech translation (Inaguma et al., 2019; Li et al., 2021) presents itself as a promising direction to alleviate data scarcity by leveraging commonalities across languages. In this multilingual translation track, we submit: 1) an end-to-end system (§5.2) that directly trans- lates from speech and 2) a cascaded system (§5.1) that consists of a multilingual speech transcription module (§3) followed by a multilingual text trans- lation module (§4). Our efforts to improve the speech translation sys- tem can be categorized as follows. When training, on the source side, we augment the speech data by speed perturbation. On the target side, we apply pseudo-labeling1 by translating the ASR transcrip- tions. Furthermore, we train multilingual systems for both speech transcription and translation to al- leviate the scarcity of training data. When testing, we use different ensembling techniques to increase the diversity of output distribution and improve output quality. The main findings from our experiments are: • Multilingual training and jointly training speech transcription and translation are benefi- cial when data scarcity limits the performance of mono- or bilingual systems. • The gain in the overall speech-to-text systems also propagates to cascaded systems as a result of stronger ASR performance. • Pseudo-labeling strongly improves speech translation quality, especially in directions that are originally zero-shot. 2 Setup 2.1 Corpus Statistics Our systems are trained on the multilingual TEDx (mTEDx) speech recognition …

Content truncated to 500 words