JHU IWSLT 2023 Multilingual Speech Translation System Description

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 302–310 July 13-14, 2023 c⃝2023 Association for Computational Linguistics JHU IWSLT 2023 Multilingual Speech Translation System Description Henry Li Xinyuan1∗ Neha Verma1∗ Bismarck Bamfo Odoom1 Ujvala Pradeep1 Matthew Wiesner2 Sanjeev Khudanpur1,2 1Center for Language and Speech Processing, and 2 Human Language Technology Center of Excellence, Johns Hopkins University {xli257, nverma7, bodoom1, upradee1, wiesner, khudanpur}@jhu.edu Abstract We describe the Johns Hopkins ACL 60-60 Speech Translation systems submitted to the IWSLT 2023 Multilingual track, where we were tasked to translate ACL presentations from English into 10 languages. We developed cascaded speech translation systems for both the constrained and unconstrained subtracks. Our systems make use of pre-trained models as well as domain-specific corpora for this highly technical evaluation-only task. We find that the specific technical domain which ACL presenta- tions fall into presents a unique challenge for both ASR and MT, and we present an error anal- ysis and an ACL-specific corpus we produced to enable further work in this area. 1 Introduction In this work, we describe the 2023 JHU 60-60 Mul- tilingual speech translation track submissions and their development (Agarwal et al., 2023; Salesky et al., 2023). This multilingual task involved the translation of ACL conference oral presentations, given in English, into 10 different target languages. High quality translation systems that can assist in translating highly technical and scientific informa- tion helps in the dissemination of knowledge to more people, which in turn can help make our field more inclusive and accessible. We briefly describe the task in Section 2. In Sec- tion 3 we describe the collection and preparation of in-domain ACL data to improve ASR and MT performance by addressing the domain-specificity of the task. We then describe our systems in Sec- tion 4, including their motivation and design in context of this shared task. Technical details of our experiments are in 5. We present our results and a discussion of our contributions in Section 6. * Authors contributed equally 2 The Speech Translation of Talks Task In 2022, the ACL began the 60-60 initiative, a di- versity and inclusion initiative to translate the ACL Anthology into 60 languages for its 60th anniver- sary. The initiative provided evaluation data for the IWSLT 2023 multilingual track on speech transla- tion of talks from English into 10 major languages. It was further split into constrained and uncon- strained subtracks. The constrained subtrack al- lowed the use of only certain datasets and pre- trained models, whereas the unconstrained subtrack had no such restrictions. We submitted systems to both subtracks and describe them in Section 4. 2.1 Evaluation Data The ACL 60-60 development data provided to par- ticipants is composed of the audio of 5 talks, their transcripts, and multi-parallel translations into 10 languages. Each talk is about 12 minutes in length – a total of about an hour of English speech for the entire set. Additionally, participants are provided with the text abstract of each talk taken from the …

Content truncated to 500 words