KIT’s IWSLT 2021 Offline Speech Translation System

Proceedings of the 18th International Conference on Spoken Language Translation, pages 125–130 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 125 KIT’s IWSLT 2021 Offline Speech Translation System Tuan-Nam Nguyen, Thai-Son Nguyen, Christian Huber, Maximilian Awiszus, Ngoc-Quan Pham, Thanh-Le Ha, Felix Schneider, Sebastian St¨uker, Alexander Waibel Karlsruhe Institute of Technology firstname.lastname@kit.edu Abstract This paper describes KIT’submission to the IWSLT 2021 Offline Speech Translation Task. We describe a system in both cascaded con- dition and end-to-end condition. In the cas- caded condition, we investigated different end- to-end architectures for the speech recognition module. For the text segmentation module, we trained a small transformer-based model on high-quality monolingual data. For the trans- lation module, our last year’s neural machine translation model was reused. In the end-to- end condition, we improved our Speech Rela- tive Transformer architecture to reach or even surpass the result of the cascade system. 1 Introduction As in previous years, the cascade system’s pipeline is constituted by an ASR module, a text segmen- tation module and a machine translation module. In this year’s evaluation campaign, we investigated only sequence-to-sequence ASR models with three architectures. The segmentation module is basi- cally a monolingual system which translates a dis- fluent, broken, uncased text (i.e. ASR outputs) into a more fluent, written-style text with punctuations in order to match the data conditions of the trans- lation system. The machine translation module’s architecture is the same as the previous year’s. For the end-to-end system, we improved from our last year’s Speech Relative Transformer architecture (Pham et al., 2020a). As a result, the end-to-end system can produce better results on certain test sets and approach the performance on some others compared to the cascade system this year, while the end-to-end system was the dominant approach last year. The rest of the paper is organized as followed. Section 2 describes the data set used to train and test the system. It is then followed by Section 3 providing the description and experimental results of both the cascade and the end-to-end system. In the end, we conclude the paper with Section 4. 2 Data Speech Corpora. For training and evaluation of our ASR models, we used Mozilla Common Voice v6.1 (Ardila et al., 2019), Europarl (Koehn, 2005), How2 (Sanabria et al., 2018), Librispeech (Panayotov et al., 2015), MuST-C v1 (Di Gangi et al., 2019), MuST-C v2 (Cattoni et al., 2021) and Tedlium v3 (Hernandez et al., 2018) dataset. The data split is presented in the following table 1. Table 1: Summary of the English data-sets used for speech recognition Corpus Utterances Speech data [h] A: Training Data Mozilla Common Voice 1225k 1667 Europarl 33k 85 How2 217k 356 Librispeech 281k 963 MuST-C v1 230k 407 MuST-C v2 251k 482 Tedlium 268k 482 B: Test Data Tedlium 1155 2.6 Librispeech 2620 5.4 Text Corpora. We collected the text parallel training data as presented in Table 2. 3 Offline Speech Translation We address the offline speech translation task by two main approaches, namely cascade and …

Content truncated to 500 words