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

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 239 - 246 May 26-27, 2022 c⃝2022 Association for Computational Linguistics The HW-TSC’s Offline Speech Translation System for IWSLT 2022 Evaluation Yinglu Li1, Minghan Wang1, Jiaxin Guo1, Xiaosong Qiao1, Yuxia Wang2, Daimeng Wei1, Chang Su1, Yimeng Chen1, Min Zhang1, Shimin Tao1, Hao Yang1, Ying Qin1 1Huawei Translation Services Center, Beijing, China 2The University of Melbourne, Melbourne, Australia {liyinglu,wangminghan,guojiaxin1,qiaoxiaosong,weidaimeng,suchang8 ,chenyimeng,zhangmin186,taoshimin,yanghao30,qinying}@huawei.com yuxiaw@student.unimelb.edu.au Abstract This paper describes the HW-TSC’s designa- tion of the Offline Speech Translation System submitted for IWSLT 2022 Evaluation. We ex- plored both cascade and end-to-end system on three language tracks (en-de, en-zh and en-ja), and we chose the cascade one as our primary submission. For the automatic speech recog- nition (ASR) model of cascade system, there are three ASR models including Conformer, S2T-Transformer and U2 trained on the mix- ture of five datasets. During inference, tran- scripts are generated with the help of domain controlled generation strategy. Context-aware reranking and ensemble based robustness en- hancement strategy are proposed to produce better ASR outputs. For machine translation part, we pretrained three translation models on WMT21 dataset and fine-tuned them on in- domain corpora. Our cascade system shows more competitive performance than the known offline systems in the industry and academia. 1 Introduction In recent years, end-to-end system and cas- cade system are fundamental pipelines for speech translation tasks. Traditional cascade system is comprised of continuing parts, automatic speech recognition (ASR) is responsible for generat- ing transcripts from audios and machine trans- lation model aims at translating ASR outputs from source language into target language. Ob- viously, the ASR part and MT part of this sys- tem are independent to some extent. Therefore, this paradigm enables people to utilise state-of- the-art ASR models and MT models and conduct experiments by different permutations and com- binations. And those experiments can help us find the best combination of choice of ASR and MT model. ASR model like Conformer (Gu- lati et al., 2020) and S2T-Transformer (Synnaeve et al., 2019) are commonly used. MT models like Transformer (Vaswani et al., 2017) can be consid- ered as a standard configuration. On the contrary, there is also a disadvantage when applying cascade systems. The main aspect is that some important information such as the in- tonation and emphasis of speakers could not be ex- plicitly expressed in the transcripts. This "missing information" might be the key to distinguish the gender of speaker, or the sarcasm and symbolism behind the texts. It means, there is a risk of los- ing important information under the condition of cascade system. Correspondingly, end-to-end system preserves the competitive edge to learn the "missing in- formation", because it is directly trained on the speech-to-text dataset without any transit process. Due to this property, end-to-end system has been paid attention in research and there is encouraging progress. For instance, Conformer (Gulati et al., 2020) can also be used in this task. However, there are some disadvantages for the end-to-end system. …

Content truncated to 500 words