DiDi Labs’ End-to-end System for the IWSLT 2020 Offline Speech TranslationTask

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 69–72 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 69 DiDi Labs’ End-to-End System for the IWSLT 2020 Offline Speech Translation Task Arkady Arkhangorodsky, Yiqi Huang, Amittai Axelrod DiDi Labs 4640 Admiralty Way Marina del Rey, CA 90292 {arkadyarkhangorodsky, yiqihuang, amittai}@didiglobal.com Abstract We describe the DiDi Labs system submitted for the IWSLT 2020 Offline Speech Transla- tion Task (Ansari et al., 2020). We trained an end-to-end system that translates audio from English TED talks to German text, without producing intermediate English text. Our base system used the S-Transformer architecture (Di Gangi et al., 2019b), trained using the MuST-C dataset (Di Gangi et al., 2019a). We extended the system via decoder pre-training, pre-trained speech features, and text transla- tion, but these extensions did not yield im- proved results. 1 Introduction The performance of end-to-end speech translation systems at IWSLT has been approaching that of cascaded systems, with the gap shrinking to 1.5 BLEU points in 2019 (Niehues et al., 2019). With additional effort, end-to-end systems could finally surpass cascaded systems. The 2020 task required participants to translate audio from English TED talks to German text. We trained several different end-to-end speech translation systems. We used the MuST-C dataset to train models for speech translation and speech recognition, the Europarl-ST dataset for speech recognition (Iranzo-S´anchez et al., 2019), and the WMT-19 news commentary dataset for text trans- lation (Tiedemann, 2012). Our best performing model used an encoder that was first pre-trained for English speech recognition, and then fine- tuned for speech translation. This system scored 17.1 BLEU on the MuST-C test set. 2 Experimental Framework Our models used the S-Transformer architecture of Di Gangi et al.. This is an adaptation of the Transformer architecture (Vaswani et al., 2017) for speech inputs. The encoder performs a 2- D convolution on the audio input before apply- ing self-attention as in the Transformer. An- other distinction is that the decoder operates at the character level, instead on the byte-pair encod- ing (BPE) tokenization that is typically used with transformer models for text. The system uses a 512-dimensional embedding in the self-attention layers. Each of the encoder and decoder have 8- headed attention and 6 self-attention layers. The models have 32,132,040 parameters. Each of our models were run on a single Nvidia Tesla P-100 GPU. We used a batch size of 8, and the Adam optimizer with a learning rate of 0.005 and an inverse square root warm-up schedule start- ing from 0.0003 for the first 4000 training steps. Each model was trained for up to 50 epochs, stop- ping early when validation loss had not decreased for 10 consecutive epochs. We trained 6 models using different methods. We used the German transcripts and German audio from Europarl-ST for decoder pre-training. We used the WMT News Commentary parallel corpus for text translation. All other experiments used the MuST-C dataset. Table 1 contains the statistics for the corpora we used. 3 Extending S-Transformer 3.1 …

Content truncated to 500 words