Efficient yet Competitive Speech Translation: FBK@IWSLT2022

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 177 - 189 May 26-27, 2022 c⃝2022 Association for Computational Linguistics Efficient yet Competitive Speech Translation: FBK@IWSLT2022 Marco Gaido1,2 , Sara Papi1,2 , Dennis Fucci1,2, Giuseppe Fiameni3, Matteo Negri1, Marco Turchi1 1Fondazione Bruno Kessler 2University of Trento 3NVIDIA AI Technology Center {mgaido, spapi, dfucci, negri, turchi}@fbk.eu Abstract The primary goal of this FBK’s systems sub- mission to the IWSLT 2022 offline and simul- taneous speech translation tasks is to reduce model training costs without sacrificing trans- lation quality. As such, we first question the need of ASR pre-training, showing that it is not essential to achieve competitive results. Sec- ond, we focus on data filtering, showing that a simple method that looks at the ratio between source and target characters yields a quality im- provement of 1 BLEU. Third, we compare dif- ferent methods to reduce the detrimental effect of the audio segmentation mismatch between training data manually segmented at sentence level and inference data that is automatically segmented. Towards the same goal of training cost reduction, we participate in the simulta- neous task with the same model trained for offline ST. The effectiveness of our lightweight training strategy is shown by the high score obtained on the MuST-C en-de corpus (26.7 BLEU) and is confirmed in high-resource data conditions by a 1.6 BLEU improvement on the IWSLT2020 test set over last year’s winning system. 1 Introduction The yearly IWSLT offline speech translation (ST) evaluation campaign aims at comparing the models produced by companies, universities, and research institutions on the task of automatically translating speech in one language into text in another lan- guage. Given a blind test set, participants’ submis- sions are ranked according to the obtained Sacre- BLEU score (Post, 2018). Over the years, the competition to achieve the highest score has driven to bigger and bigger mod- els trained on large datasets: the 2021 winning model (Bahar et al., 2021b) has twice the number of encoder layers (12 vs 6), and a deeper (6 vs 4 layers) and larger (1024 vs 512 features) decoder The authors contributed equally. compared to the 2019 winner (Potapczyk et al., 2019). In addition, most of the competitors have relied on knowledge transfer techniques (Ansari et al., 2020; Anastasopoulos et al., 2021b), such as the initialization of the ST model encoder with the encoder of an ASR system trained on large cor- pora (Bansal et al., 2019). All these practices have contributed to a remarkable increase in computa- tional expenses and energy consumption that are antithetic with the recent rise of concerns on the social and environmental consequences of these costs (Strubell et al., 2019). Among the harms inherent to the high computa- tional cost of training ST systems, there is also the risk of restricting the participation in competitions like IWSLT to few big players from the industry sectors that can afford them. As part of a research institution, with this work we try to answer the question: …

Content truncated to 500 words