QUESPA Submission for the IWSLT 2023 Dialect and Low-resource Speech Translation Tasks

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 261–268 July 13-14, 2023 c⃝2023 Association for Computational Linguistics QUESPA Submission for the IWSLT 2023 Dialect and Low-resource Speech Translation Tasks John E. Ortega1, Rodolfo Zevallos2, and William Chen3 1Northeastern University, USA, 2Universitat de Pompeu Fabra, Spain 3Carnegie Mellon University, USA contact email: j.ortega@northeastern.edu Abstract This article describes the QUESPA team speech translation (ST) submissions for the Quechua to Spanish (QUE–SPA) track featured in the Evaluation Campaign of IWSLT 2023: low- resource and dialect speech translation. Two main submission types were supported in the campaign: constrained and unconstrained. We submitted six total systems of which our best (primary) constrained system consisted of an ST model based on the Fairseq S2T framework where the audio representations were created using log mel-scale filter banks as features and the translations were performed using a trans- former. The best (primary) unconstrained sys- tem used a pipeline approach which combined automatic speech recognition (ASR) with ma- chine translation (MT). The ASR transcriptions for the best unconstrained system were com- puted using a pre-trained XLS-R–based model along with a fine-tuned language model. Tran- scriptions were translated using a MT system based on a fine-tuned, pre-trained language model (PLM). The four other submissions are presented in this article (2 constrained and 2 unconstrained) for comparison because they consist of various architectures. Our results show that direct ST (ASR and MT combined together) can be more effective than a PLM in a low-resource (constrained) setting for Quechua to Spanish. On the other hand, we show that fine-tuning of any type on both the ASR and MT system is worthwhile, resulting in nearly 16 BLEU for the unconstrained task. 1 Introduction Low-resource machine translation (LRMT) can be considered a difficult task due to the low amount of parallel data on hand. (Haddow et al., 2022) By adding the task of automatic speech recogni- tion (ASR), complexity can be even more difficult. Findings from the previous year’s IWSLT 2022 (Antonios et al., 2022) have shown that for low- resource language pairs like Tamasheq–French, it is difficult to achieve more than 5 BLEU (Papineni et al., 2002) score points for the combined task of speech translation (ST), even in a unconstrained setting. This year, the IWSLT 2023 (Agarwal et al., 2023) evaluation campaign for low-resource and dialect speech translation has included Tamasheq– French along with several other language pairs. One of the newly introduced language pairs is Quechua–Spanish deemed QUE–SPA by the orga- nizers. Quechua is an indigenous language spoken in the Andes mountainous region in South America. It is spoken by millions of native speakers mostly from Peru, Ecuador and Bolivia. In those regions, the high-resource language is Spanish. Quechua displays many unique morphological properties of which high inflection and poly-synthetic are the two most commonly known. It is worthwhile to note that previous work (Ortega and Pillaipakkam- natt, 2018; Ortega et al., 2020) has been somewhat successful in identifying the inflectional properties of Quechua such as …

Content truncated to 500 words