Samsung’s System for the IWSLT 2019 End-to-End Speech Translation Task

Samsung’s System for the IWSLT 2019 End-to-End Speech Translation Task Tomasz Potapczyk, Pawel Przybysz, Marcin Chochowski, Artur Szumaczuk Samsung R&D Institute, Poland {t.potapczyk,p.przybysz,m.chochowski,a.szumaczuk}@samsung.com Abstract This paper describes the submission to IWSLT 2019 End- to-End speech translation task by Samsung R&D Institute, Poland. We decided to focus on end-to-end English to German TED lectures translation and did not provide any submission for other speech tasks. We used a slightly al- tered Transformer [1] architecture with standard convolu- tional layer preparing the audio input to Transformer en- coder. Additionally, we propose an audio segmentation al- gorithm maximizing BLEU score on tst2015 test set. 1. Introduction This paper describes the submission to IWSLT 2019 End- to-End Speech Translation task by Samsung R&D Institute, Poland. System architecture and data preparation techniques were designed before IWSLT 2019 data were released. We have been using LibriSpeech corpus [2] to develop these techniques. We evaluated our models on TED 2010 test set. After IWSLT 2019 training data was published we gathered successful techniques to train a systems for this competition. These techniques were used in all three models that will be presented here. Document structure is as follows. Firstly we describe data preparation and augmentation. Then we provide system specification and training procedure used in our experiments. We describe data segmentation algorithm used to segment test sets TED 2015 and TED 2019. We show results of our experiments with monolingual data and simple recurrent unit (SRU) [3]. Finally we conclude our results. 2. Training Data To train our system we used only IWSLT 2019 permissible audio corpora - iwslt-corpus, TEDLIUM2[4] and MUST-C corpus[5]. Data preparation process started with data filter- ing. Then we generated synthetic target sentences with text- to-text machine translation model. We augmented audio in- put with sox1. Finally we included monolingual text data with empty audio input. 1SoX - Sound eXchange sox.sourceforge.net v14.4.1 2.1. Data filtration We trained English ASR system that was used to filter iwslt- corpus and TEDLIUM2 corpora. We removed cases where WER score exceeded 75% when comparing ASR output and English reference. We decided that MUST-C corpus does not need filtration. Additionally we filtered iwslt-corpus with regard to quality of translation using statistical dictionary- based methods. Size of the corpora before and after filtration is shown in Table 1. Corpora Orig. size Filtered Length iwslt-corpus (ASR) 171121 158737 224h + trans. quality 158737 126817 188h TEDLIUM2 92973 90715 197h MUST-C 229703 229703 400h Table 1: Size (number of audio utterances) of the training corpora before and after filtration. Iwslt-corpus (ASR) is corpus filtered by ASR only. The last column is total audio length after filtration. 2.2. Synthetic target data TEDLIUM2 corpus did not provide any German translations, therefore we generated synthetic targets using two Trans- former Big MT systems trained with different hyperparame- ters on WMT data - Paracrawl, Europarl and OpenSubtitles. Training data for these systems has been prepared with our in-house data preparation pipeline. We also used synthetic translations as an alternative translation in iwslt-corpus when augmenting it. …

Content truncated to 500 words