KIT’s IWSLT 2020 SLT Translation System

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 55–61 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 55 KIT’s IWSLT 2020 SLT Translation System Ngoc-Quan Pham, Felix Schneider, Tuan-Nam Nguyen, Thanh-Le Ha, Thai-Son Nguyen, Maximilian Awiszus, Sebastian St¨uker, Alexander Waibel Karlsruhe Institute of Technology firstname.lastname@kit.edu Abstract This paper describes KIT’s submissions to the IWSLT2020 Speech Translation evaluation campaign. We first participate in the simulta- neous translation task, in which our simultane- ous models are Transformer-based and can be efficiently trained to obtain low latency with minimized compromise in quality. On the of- fline speech translation task, we applied our new Speech Transformer architecture to end- to-end speech translation. The obtained model can provide translation quality which is com- petitive to a complicated cascade. The latter still has the upper hand, thanks to the ability to transparently access to the transcription, and resegment the inputs to avoid fragmentation. 1 Introduction The Karlsruhe Institute of Technology (KIT) par- ticipated in the IWSLT 2020 Evaluation Cam- paign (Ansari et al., 2020) in two main tracks: Of- fline Speech Translation task (SLT) and Simulta- neous Text Translation. Our highlight s the pro- posal of a novel method for training simultaneous translation models, with the Adaptive Computa- tion Time technique (Graves, 2016) incorporated to the Transformer models (Vaswani et al., 2017). On the other hand, the end-to-end speech transla- tion models have observed a single deep Speech Transformer (Pham et al., 2019b) approaching the performance of a heavily powered cascade. The latter, however, is more transparent because of vis- ible inputs and outputs to each components. It is still the dominant approach, thanks to the segmen- tation module that adds punctuations and sentence boundaries, so the MT models do not suffer from fragmentation. 2 Data The overall data that the project employed can be divided into two main sections: speech and text corpora. Speech Corpora. We gathered the allowed train- ing data included MuST-C and Speech-Translation TED Talks containing both parallel data for au- dio to English and German. The TEDLIUM3 and the Mozilla Common Voice data are speech recognition-specific. Furthermore we also consid- ered the How2 dataset (the Portuguese translation is ignored). The data is further cleaned with ASR models (the details are unveiled in Section 4.4) to obtain the training time as shown in Table 1. Table 1: Speech Training data Data Segments Total time MuST-C 229K 408h Speech Translation 142K 160h TEDLIUM 264K 415h Common Voice 854K 1490h How2 217K 360h Text Corpora. We collected the text parallel training data as presented in Table 2. Table 2: Text Training Data Dataset Sentences TED Talks (TED) 220K Europarl (EPPS) 2.2MK CommonCrawl 2.1M Rapid 1.21M ParaCrawl 25.1M OpenSubtitles 12.6M WikiTitle 423K Back-translated News 26M 56 3 Simultaneous Speech Translation For simultaneous speech translation, we deploy a novel model on the text-to-text task based on Adaptive Computation Time (ACT, Graves (2016)). At each decoder step, the model makes a decision on whether to READ another input token or to WRITE …

Content truncated to 500 words