The USTC-NELSLIP Offline Speech Translation Systems for IWSLT 2022

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 198 - 207 May 26-27, 2022 c⃝2022 Association for Computational Linguistics The USTC-NELSLIP Offline Speech Translation Systems for IWSLT 2022 Weitai Zhang1,2, Zhongyi Ye2, Haitao Tang2, Xiaoxi Li2, Xinyuan Zhou2, Jing Yang2, Jianwei Cui1, Pan Deng1, Mohan Shi1, Yifan Song1, Dan Liu1,2, Junhua Liu1,2, and Lirong Dai1 1University of Science and Technology of China, Hefei, China 2iFlytek Research, Hefei, China {zwt2021, danliu, jwcui, pdeng, smohan, yfsong}@mail.ustc.edu.cn lrdai@ustc.edu.cn {zyye7, xxli16, httang, xyzhou15, jingyang24, jhliu}@iflytek.com Abstract This paper describes USTC-NELSLIP’s sub- missions to the IWSLT 2022 Offline Speech Translation task, including speech translation of talks from English to German, English to Chinese and English to Japanese. We de- scribe both cascaded architectures and end-to- end models which can directly translate source speech into target text. In the cascaded con- dition, we investigate the effectiveness of dif- ferent model architectures with robust training and achieve 2.72 BLEU improvements over last year’s optimal system on MuST-C English- German test set. In the end-to-end condition, we build models based on Transformer and Conformer architectures, achieving 2.26 BLEU improvements over last year’s optimal end-to- end system. The end-to-end system has ob- tained promising results, but it is still lagging behind our cascaded models. 1 Introduction This paper describes the submission to IWSLT 2022 Offline Speech Translation task by National Engineering Laboratory for Speech and Language Information Processing (NELSLIP), University of Science and Technology of China, China. For years, Spoken Language Translation (SLT) has been addressed by cascading an Automatic Speech Recognition (ASR) and a Machine Trans- lation (MT) system. The ASR system processes source speech into source text and the MT sys- tem translates ASR output into text in target lan- guage independently. Recent trends rely on using a single neural network to directly translate the speech in source language into the text in target language without intermediate symbolic represen- tations. The end-to-end paradigm shows an enor- mous potential to overcome some of the cascaded systems’ problems, such as higher architectural complexity and error propagation (Duong et al., 2016; Berard et al., 2016; Weiss et al., 2017). Last year’s results of IWSLT 2021 have confirmed that the performance of end-to-end models is approach- ing the results of cascaded solutions. The best end-to-end submission (under the same segmenta- tion and training data conditions) is 2 BLEU points (22.6 vs 24.6) below the top-ranked system (Anas- tasopoulos et al., 2021). In this work, we build machine translation sys- tems with techniques like back translation (Sen- nrich et al., 2016a), domain adaptation and model ensemble, which have been proved to be effec- tive practices in IWSLT and WMT (Akhbardeh et al., 2021). Besides, we further improve cas- caded speech translation system performance with methods of self-training (Kim and Rush, 2016; Ren et al., 2020; Liu et al., 2019), speech synthesis (Shen et al., 2018), Supervised Hybrid Audio Seg- mentation (SHAS) (Tsiamas et al., 2022), etc. In end-to-end condition, we initialize the en- coder with the corresponding …

Content truncated to 500 words