UM-DFKI Maltese Speech Translation

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 433–441 July 13-14, 2023 c⃝2023 Association for Computational Linguistics UM-DFKI Maltese Speech Translation Aiden Williams* Kurt Abela* Rishu Kumar⋄ Martin Bär* Hannah Billinghurst* Kurt Micallef* Ahnaf Mozib Samin* Andrea De Marco* Lonneke van der Plas† Claudia Borg* *University of Malta, ⋄DFKI, †IDIAP aiden.williams.19@um.edu.mt, kurt.abela@um.edu.mt, rishu.Kumar@dfki.de,martin.bar.22@um.edu.mt, hannah.billinghurst.22@um.edu.mt, kurt.micallef@um.edu.mt, ahnaf.samin.22@um.edu.mt, andrea.demarco@um.edu.mt, lonneke.vanderplas@idiap.ch, claudia.borg@um.edu.mt Abstract For the 2023 IWSLT (Agarwal et al., 2023) Maltese Speech Translation Task, UM-DFKI jointly presents a cascade solution which achieves 0.6 BLEU. While this is the first time that a Maltese speech translation task has been released by IWSLT, this paper explores pre- vious solutions for other speech translation tasks, focusing primarily on low-resource sce- narios. Moreover, we present our method of fine-tuning XLS-R models for Maltese ASR us- ing a collection of multi-lingual speech corpora as well as the fine-tuning of the mBART model for Maltese to English machine translation. 1 Introduction Speech Translation (ST), or speech-to-text trans- lation, involves converting speech in a source lan- guage into written text in a target language. With the rise of deep learning, steep progress has been made in this field and many other areas that fall under the Natural Language Processing (NLP) um- brella (Khurana et al., 2023; Qiu et al., 2020). How- ever, development for low-resource languages has continued to present difficulties and obstacles due to a variety of factors, including the lack of suffi- cient training data, language experts and other re- sources (Magueresse et al., 2020; Hedderich et al., 2021). The International Workshop on Spoken Lan- guage Translation (IWSLT) shared task is an an- nual competition that aims to foster research in the field of speech translation. With its low-resource track, it also contributes to advanced research for speech translation in low-resource scenarios. In this paper, we present our submission to the low re- source track: a pipeline system for English-Maltese speech-to-text translation. We begin by discussing the state of the art in speech translation and describe the two main ap- proaches, cascade and end-to-end. Afterwards, we briefly summarise the challenges posed by low- resource languages and possible mitigation strate- gies. We then describe our system, a pipeline ap- proach containing an internal Automatic Speech Recognition (ASR) component and the outward facing Machine Translation (MT) component. The ASR component can use one of five fine-tuned XLS-R (Babu et al., 2021) models, whereas the MT stage always uses an mBART-50 model. 2 Literature Review The following literature review aims to provide an overview of previous IWSLT ST submissions, with a particular focus on low-resource scenarios. The review is divided into two sections; where the first explores the general approaches and challenges as- sociated with low-resource ST, and the second sec- tion discusses previous approaches to low-resource ST as applied to IWSLT. 2.1 Previous IWSLT Approaches for Low-Resource Languages The IWSLT (Anastasopoulos et al., 2022) set the task in 2022 to attempt to solve “the problem of de- veloping speech transcription and translation tools for …

Content truncated to 500 words