The Sogou-TIIC Speech Translation System for IWSLT 2018

The Sogou-TIIC Speech Translation System for IWSLT 2018 Yuguang Wang*, Liangliang Shi*, Linyu Wei*, Weifeng Zhu*, Jinkun Chen* Zhichao Wang*, Shixue Wen*, Wei Chen*, Yanfeng Wang*, Jia Jia† *Voice Interaction Technology Center, Sogou Inc., Beijing, China {wangyuguang,shiliangliang}@sogou-inc.com †Tiangong Institute for Intelligent Computing, Tsinghua University, Beijing, China {jjia}@tsinghua.edu.cn Abstract This paper describes our speech translation system for the IWSLT 2018 Speech Translation of lectures and TED talks from English to German task. The pipeline approach is employed in our work, which mainly includes the Automatic Speech Recognition (ASR) system, a post-processing module, and the Neural Machine Translation (NMT) system. Our ASR system is an ensemble system of Deep-CNN, BLSTM, TDNN, N-gram Language model with lattice rescoring. We report average results on tst2013, tst2014, tst2015. Our best combination system has an average WER of 6.73. The machine translation system is based on Google’s Transformer architecture. We achieved an improvement of 3.6 BLEU over baseline system by applying several techniques, such as cleaning parallel corpus, fine tuning of single model, ensemble models and re-scoring with additional features. Our final average result on speech translation is 31.02 BLEU. 1. Introduction We have participated in the Speech Translation of lectures and TED talks from English to German task. The goal of this task is to translate fully un-segmented talks or lectures from English to German. A pipeline approach is employed in our work. It consists of segmentation of audio data, ASR system, punctuation restoration and NMT system. A two pass decoding is used in the ASR system. In the first pass, we use several different neural network, such as Deep-CNN [1] [2], BLSTM and TDNN [3] to generate ensemble results. Then the decoding lattices of the ensemble system are sent to a second pass decoder for lattice rescoring. In order to bridge the gap between the output of ASR system and training data of NMT system, punctuation restoration, disfluency detection and inverse text normalization are necessary in our pipeline. Our NMT system is based on the Transformer architecture [4], which is based solely on attention mechanisms. Several techniques are adopted to improve our system, such as parallel corpus cleaning, fine tuning, model ensembling and re-scoring with additional features. The rest of this paper is structured as follows. Section 2 describes the details of our ASR system, and Section 3 describes our NMT system. Our results in the speech translation task are presented in Section 4. We conclude this paper in Section 5. 1https://github.com/kaldi- asr/kaldi/blob/master/egs/ami/s5b/local/run_cleanup _segmentation.sh 2. Automatic speech recognition 2.1. Audio Segmentation In this evaluation, the test set is provided without manual sentence segmentation, thus automatic segmentation of the final test set is essential. We utilize an approach to automatic segment audio data based on the signal energy. We set a threshold to split the audio between 8 and 15 seconds and then concatenate utterances that are shorter than 8 seconds to its neighboring utterances. 2.2. Audio Data Preparation and Feature Extraction 2.2.1. Data Cleaning Our acoustic data comes from two sources. The first is …

Content truncated to 500 words