ON-TRAC Consortium Systems for the IWSLT 2022 Dialect and Low-resource Speech Translation Tasks

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 308 - 318 May 26-27, 2022 c⃝2022 Association for Computational Linguistics ON-TRAC Consortium Systems for the IWSLT 2022 Dialect and Low-resource Speech Translation Tasks Marcely Zanon Boito1, John Ortega2, Hugo Riguidel2, Antoine Laurent2, Loïc Barrault2, Fethi Bougares3, Firas Chaabani3, Ha Nguyen1,5, Florentin Barbier4, Souhir Gahbiche4, Yannick Estève1 1LIA - Avignon University, France, 2LIUM - Le Mans University, France 3ELYADATA - Tunis, Tunisia, 4Airbus - France, 5LIG - Grenoble Alpes University contact email: yannick.esteve at univ-avignon.fr Abstract This paper describes the ON-TRAC Consor- tium translation systems developed for two challenge tracks featured in the Evaluation Campaign of IWSLT 2022: low-resource and dialect speech translation. For the Tunisian Arabic-English dataset (low-resource and di- alect tracks), we build an end-to-end model as our joint primary submission, and com- pare it against cascaded models that lever- age a large fine-tuned wav2vec 2.0 model for ASR. Our results show that in our set- tings pipeline approaches are still very com- petitive, and that with the use of transfer learn- ing, they can outperform end-to-end models for speech translation (ST). For the Tamasheq- French dataset (low-resource track) our primary submission leverages intermediate representa- tions from a wav2vec 2.0 model trained on 234 hours of Tamasheq audio, while our con- trastive model uses a French phonetic transcrip- tion of the Tamasheq audio as input in a Con- former speech translation architecture jointly trained on automatic speech recognition, ST and machine translation losses. Our results highlight that self-supervised models trained on smaller sets of target data are more effec- tive to low-resource end-to-end ST fine-tuning, compared to large off-the-shelf models. Results also illustrate that even approximate phonetic transcriptions can improve ST scores. 1 Introduction The vast majority of speech pipelines are devel- oped for and in high-resource languages, a small percentage of languages for which there is a large amount of annotated data freely available (Joshi et al., 2020). However, the assessment of systems’ performance only on high-resource settings can be problematic because it fails to reflect the real-world performance these approaches will have in diverse and smaller datasets. In this context, the IWSLT 2022 (Anastasopou- los et al., 2022) proposes two interesting shared tasks: low-resource and dialect speech transla- tion (ST). The former aims to assess the exploitabil- ity of current translation systems in data scarcity settings. The latter focuses on the assessment of the systems capabilities in noisy settings: differ- ent dialects are mixed in a single dataset of spon- taneous speech. For the low-resource task, this year’s language pairs are: Tamasheq-French and Tunisian Arabic-English. The latter is also used, in constrained conditions, for the dialect task. This paper reports the ON-TRAC consortium submissions for the mentioned tasks. The ON- TRAC Consortium is composed of researchers from three French academic laboratories, LIA (Avi- gnon University), LIUM (Le Mans University) and LIG (University Grenoble Alpes), together with two industrial partners: Airbus France and ELY- DATA. Our systems for the dialect task focus on the …

Content truncated to 500 words