Without Further Ado: Direct and Simultaneous Speech Translation by AppTek in 2021

Proceedings of the 18th International Conference on Spoken Language Translation, pages 52–63 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 52 Without Further Ado: Direct and Simultaneous Speech Translation by AppTek in 2021 Parnia Bahar∗, Patrick Wilken∗, Mattia di Gangi, Evgeny Matusov Applications Technology (AppTek), Aachen, Germany {pbahar,pwilken,mdigangi,ematusov}@apptek.com Abstract This paper describes the offline and simulta- neous speech translation (ST) systems devel- oped at AppTek for IWSLT 2021. Our of- fline ST submission includes the direct end- to-end system and the so-called posterior tight integrated model, which is akin to the cas- cade system but is trained in an end-to-end fashion, where all the cascaded modules are end-to-end models themselves. For simultane- ous ST, we combine hybrid automatic speech recognition (ASR) with a machine translation (MT) approach whose translation policy deci- sions are learned from statistical word align- ments. Compared to last year, we improve general quality and provide a wider range of quality/latency trade-offs, both due to a data augmentation method making the MT model robust to varying chunk sizes. Finally, we present a method for ASR output segmentation into sentences that introduces a minimal addi- tional delay. 1 Introduction In this paper, we describe the AppTek speech trans- lation systems that participate in the offline and simultaneous tracks of the IWSLT 2021 evaluation campaign. This paper is organized as follows: In Section 2, we briefly address our data preparation. Section 3 describes our offline ST models followed by the experimental results in Section 3.6. For the offline end-to-end translation task, we train deep Transformer models that benefit from pretraining, data augmentation in the form of synthetic data and SpecAugment, as well as domain adaptation on TED talks. Motivated by Bahar et al. (2021), we also collapse the ASR and MT components into a posterior model which passes on the ASR pos- teriors as input to the MT model. This system is not considered a direct model since it is closer to ∗equal contribution the cascade system while being end-to-end train- able. Our simultaneous translation systems are covered in Section 4 with discussions on experi- mental results in Section 4.5. We resume the work on our streaming MT model developed for IWSLT 2020, which is based on splitting the stream of in- put words into chunks learned from statistical word alignment. Most notably, we can implement a flexi- ble quality/latency trade-off by simulating different latencies at training time. We also meet this year’s requirement to support unsegmented input by de- veloping a neural sentence segmenter that splits the ASR output into suitable translation units, using a varying number of future words as context which minimizes the latency added by this component. The experiments have been done using RASR (Wiesler et al., 2014), RETURNN (Zeyer et al., 2018a), and Sisyphus (Peter et al., 2018). 2 Data Preparation 2.1 Text Data We participate in the constrained condition and divide the allowed bilingual training data into in- domain (the TED and MuST-C v2 corpora), clean (the NewsCommentary, Europarl, and …

Content truncated to 500 words