NAVER LABS Europe’s Multilingual Speech Translation Systems for the IWSLT 2023 Low-Resource Track

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 144–158 July 13-14, 2023 c⃝2023 Association for Computational Linguistics NAVER LABS Europe’s Multilingual Speech Translation Systems for the IWSLT 2023 Low-Resource Track Edward Gow-Smith1∗ Alexandre Bérard2† Marcely Zanon Boito2† Ioan Calapodescu2 1 University of Sheffield 2 NAVER LABS Europe egow-smith1@sheffield.ac.uk first.last@naverlabs.com Abstract This paper presents NAVER LABS Europe’s systems for Tamasheq-French and Quechua- Spanish speech translation in the IWSLT 2023 Low-Resource track. Our work attempts to maximize translation quality in low-resource settings using multilingual parameter-efficient solutions that leverage strong pre-trained mod- els. Our primary submission for Tamasheq outperforms the previous state of the art by 7.5 BLEU points on the IWSLT 2022 test set, and achieves 23.6 BLEU on this year’s test set, outperforming the second best participant by 7.7 points. For Quechua, we also rank first and achieve 17.7 BLEU, despite having only two hours of translation data. Finally, we show that our proposed multilingual architecture is also competitive for high-resource languages, out- performing the best unconstrained submission to the IWSLT 2021 Multilingual track, despite using much less training data and compute. 1 Introduction The vast majority of speech pipelines are developed for high-resource languages, a small percentage of languages that have ample amounts of annotated data available (Joshi et al., 2020). However, the assessment of systems’ performance based only on high-resource settings can be problematic, since it fails to reflect the real-world performance these ap- proaches will have in diverse and smaller datasets. Moreover, as around half of the world’s languages are considered to be not only low-resource, but also from oral tradition (i.e., without a written form), there is an urgent need for speech technol- ogy that can operate robustly in such low-resource settings (Bird, 2011). In this context, the IWSLT conference1 proposes low-resource speech trans- lation (ST) challenges that allow the speech com- munity to realistically benchmark ST approaches rope.∗Work done during an internship at NAVER LABS Eu- †Equal contribution 1https://iwslt.org/ using diverse and representative datasets. This paper describes NAVER LABS Europe’s (NLE) submission to two of the language pairs from the IWSLT 2023 (Agarwal et al., 2023) Low-Resource Track: Tamasheq-French (Taq-Fr) and Quechua- Spanish (Que-Es). Most successful approaches for tackling scenar- ios where ST data is scarce perform transfer learn- ing across languages and modalities, leveraging multilingual pre-trained models for both speech and text (Anastasopoulos et al., 2022). However, due to the large number of parameters of cur- rent Transformer-based (Vaswani et al., 2017) ap- proaches, training such systems is computationally expensive and not accessible to everyone. NLE’s submission focuses on a multilingual parameter- efficient training solution that allows us to lever- age strong pre-trained speech and text models to maximize performance in low-resource lan- guages. We present new SOTA results for the Taq- Fr pair (17 hours of training data) that represent a 57% BLEU increase compared to the results achieved by Khurana et al. (IWSLT 2022 post- evaluation).2 This same system achieves 23.6 BLEU on the IWSLT 2023 test set, …

Content truncated to 500 words