SRPOL’s System for the IWSLT 2020 End-to-End Speech Translation Task

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 89–94 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 89 SRPOL’s System for the IWSLT 2020 End-to-End Speech Translation Task Tomasz Potapczyk, Pawel Przybysz Samsung R&D Institute, Poland {t.potapczyk,p.przybysz}@samsung.com Abstract This paper describes the submission to IWSLT 2020(Ansari et al., 2020) End-to-End speech translation task by Samsung R&D Institute, Poland. We took part in the offline End-to-End English to German TED lectures translation task. We based our solution on our last year’s submission(Potapczyk et al., 2019). We used a slightly altered Transformer(Vaswani et al., 2017) architecture with ResNet-like(He et al., 2016) convolutional layer preparing the audio input to Transformer encoder. To improve the model’s quality of translation we introduced two regularization techniques and trained on machine translated Librispeech(Panayotov et al., 2015) corpus in addition to iwslt- corpus, TEDLIUM2(Rousseau et al., 2014) and Must C(Di Gangi et al., 2019) corpora. Our best model scored almost 3 BLEU higher than last year’s model. To segment 2020 test set we used exactly the same procedure as last year. 1 Introduction This paper describes the submission to IWSLT 2020 End-to-End Speech Translation task by Sam- sung R&D Institute, Poland. We propose a few improvements to our previous system. Introducing additional training data gave us 0.7 BLEU improvement. Spectrogram augmen- tation techniques increased quality by 0.2 BLEU. Encoder layer depth of 12 layers gives an increase of BLEU by 1 point and 1.8 points when combined with additional training data and two spectrogram augmentation techniques. Replacing simpler con- volutions with ResNet-like convolutional layers gave around 0.5 BLEU improvement. Combining all of these and increasing embedding size to 512 resulted in almost 3 BLEU improvement compared to our last year’s model. Document structure is as follows. Firstly we de- scribe data preparation and augmentation. Then we provide system specification and training pro- cedure used in our experiments. We describe data segmentation algorithm used to segment test sets TED 2019 and TED 2020. We show results of our experiments. Finally we conclude our results. 2 Training Data To train our system we used only IWSLT 2020 per- missible audio corpora - iwslt-corpus, TEDLIUM2 (Rousseau et al., 2014), Must C corpus(Di Gangi et al., 2019) and machine translated Librispeech (Panayotov et al., 2015) corpus. We did not do any further data preparation in case of iwslt-corpus and TEDLIUM2, we used the same data as in 2019. In case of Must C corpus, this year we ran a train- ing with half of translations being synthetic. This improved the score by 0.35 BLEU. We used the same text translation models as last year to gen- erate synthetic translations. Both models scored around 33.8 and 31.1 BLEU on tst2010 and tst2015 sets respectively. 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 refer- ence. We decided that Must C corpus does not need filtration. Additionally, …

Content truncated to 500 words