The HW-TSC’s Simultaneous Speech Translation System for IWSLT 2022 Evaluation

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 247 - 254 May 26-27, 2022 c⃝2022 Association for Computational Linguistics The HW-TSC’s Simultaneous Speech Translation System for IWSLT 2022 Evaluation Minghan Wang1, Jiaxin Guo1, Yinglu Li1, Xiaosong Qiao1, Yuxia Wang2, Zongyao Li1, Chang Su1, Yimeng Chen1, Min Zhang1, Shimin Tao1, Hao Yang1, Ying Qin1 1Huawei Translation Services Center, Beijing, China 2The University of Melbourne, Melbourne, Australia {wangminghan,guojiaxin1,liyinglu,qiaoxiaosong,lizongyao,suchang8 ,chenyimeng,zhangmin186,taoshimin,yanghao30,qinying}@huawei.com yuxiaw@student.unimelb.edu.au Abstract This paper presents our work in the partici- pation of IWSLT 2022 simultaneous speech translation evaluation. For the track of text- to-text (T2T), we participate in three language pairs and build wait-k based simultaneous MT (SimulMT) model for the task. The model was pretrained on WMT21 news corpora, and was further improved with in-domain fine-tuning and self-training. For the speech-to-text (S2T) track, we designed both cascade and end-to-end form in three language pairs. The cascade sys- tem is composed of a chunking-based stream- ing ASR model and the SimulMT model used in the T2T track. The end-to-end system is a simultaneous speech translation (SimulST) model based on wait-k strategy, which is di- rectly trained on a synthetic corpus produced by translating all texts of ASR corpora into spe- cific target language with an offline MT model. It also contains a heuristic sentence breaking strategy, preventing it from finishing the transla- tion before the the end of the speech. We evalu- ate our systems on the MUST-C tst-COMMON dataset and show that the end-to-end system is competitive to the cascade one. Meanwhile, we also demonstrate that the SimulMT model can be efficiently optimized by these approaches, resulting in the improvements of 1-2 BLEU points. 1 Introduction Simultaneous speech/text translation (SimulST/SimulMT) applications are widely demanded in international communication scenarios such as conferences or live streaming. From the perspective of system architecture, re- cent works on SimulST can be classified into cas- cade and end-to-end forms. Cascade systems are often composed of a streaming Automatic Speech Recognition (ASR) module and a steaming text- to-text machine translation module (MT). It might also contains other correction modules. The inte- gration of these modules can be challenging, but the training of each can be beneficial from suf- ficient data resources. End-to-end approach is also a choice for SimulST, where translations can be directly generated from a unified model with the speech inputs, but bilingual speech translation datasets are still scarce resources. From the perspective of simultaneous strategy, there is a fixed strategy which is represented by wait-k (Ma et al., 2019) and a flexible strategy such as monotonic attention (Arivazhagan et al., 2019). The fixed strategy is easier to implement but with inferior performance and the flexible one is more robust to the speed of speech but can be non-trivial in the implementation and training. Re-translation is also a strategy proposed recently for SimulMT system, which benefits from pre-trained MT mod- els but often encounters with flicker (Arivazhagan et al., 2020; Sen et al., 2021). The IWSLT 2022 SimulST shared task (Anas- …

Content truncated to 500 words