BIT’s System for Multilingual Track

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 455–460 July 13-14, 2023 c⃝2023 Association for Computational Linguistics BIT’s System for Multilingual Track Zhipeng Wang Beijing Institute of Technology wzp3139725181@163.com Yuhang Guo∗ Beijing Institute of Technology guoyuhang@bit.edu.cn Shuoying Chen Beijing Institute of Technology chensy@bit.edu.cn Abstract This paper describes the system we submitted to the IWSLT 2023 multilingual speech transla- tion track, with the input is speech from one lan- guage, and the output is text from 10 target lan- guages. Our system consists of CNN and Trans- former, convolutional neural networks down- sample speech features and extract local infor- mation, while transformer extract global fea- tures and output the final results. In our system, we use speech recognition tasks to pre-train encoder parameters, and then use speech trans- lation corpus to train the multilingual speech translation model. We have also adopted other methods to optimize the model, such as data augmentation, model ensemble, etc. Our sys- tem can obtain satisfactory results on test sets of 10 languages in the MUST-C corpus. 1 Introduction Speech translation refers to the technology of trans- lating source language speech into target language text (or speech). This task has a very broad applica- tion space in real life, such as in international con- ferences, lectures, and overseas tourism; Adding speech translation to short videos or real-time sub- titles in some foreign language videos can provide users with a better experience. Early speech trans- lation is the combination of speech recognition and machine translation. Firstly, the speech recognition model recognizes source language speech as source language transcribed text, and then the machine translation model translates the recognized source language text into the target language text, which is also called cascade method. The advantage of cascade model is that it can use a large amount of data in speech recognition and machine translation to train the model, and it is relatively simple to implement. However, the disadvantages of cascade model are also obvious: errors in speech recogni- tion results will be transferred to the next machine ∗*Corresponding author translation task. So researchers have focused on end-to-end speech translation. At present, bilingual end-to-end speech translation has achieved very good results, but using a single model to complete multiple language translations has always been a goal pursued by researchers, that is multilingual speech translation. Compared to bilingual speech translation, the advantages of multilingual speech translation include: (1) completing multilingual translation with fewer parameters; (2) low resource languages can learn knowledge from high resource languages. In this paper, we conducted one-to- many multilingual speech translation, and submit- ted our system to the IWSLT 2023(Agarwal et al., 2023) multilingual speech translation track. Here is an introduction to our submitted system: We first use convolutional neural networks to downsample the input features, then input them into the Transformer model for further processing, and finally output the translation results at the out- put layer. The encoder for speech translation needs to complete both acoustic feature extraction and se- mantic …

Content truncated to 500 words