Alibaba Speech Translation Systems for IWSLT 2018

Alibaba Speech Translation Systems for IWSLT 2018 Nguyen Bach*, Hongjie Chen*, Kai Fan*, Cheung-Chi Leung*, Bo Li*, Chongjia Ni*, Rong Tong*, Pei Zhang*, Boxing Chen, Bin Ma, Fei Huang Machine Intelligence Technology Lab, Alibaba Group firstname.lastname@alibaba-inc.org Abstract This work describes the En→De Alibaba speech translation system developed for the evaluation campaign of the Interna- tional Workshop on Spoken Language Translation (IWSLT) 2018. In order to improve ASR performance, multiple ASR models including conventional and end-to-end models are built, then we apply model fusion in the final step. ASR pre- and post- processing techniques such as speech segmenta- tion, punctuation insertion, and sentence splitting are found to be very useful for MT. We also employed most techniques that have proven effective during the WMT 2018 evaluation, such as BPE, back translation, data selection, model ensem- bling and reranking. These ASR and MT techniques, com- bined, improve the speech translation quality significantly. 1. Introduction In this paper we describe the Alibaba speech translation sys- tem that was built as part of the Speech Translation Task in IWSLT 2018. The task involved translating English audio to German text in which English audio are from lectures and TED talks. Our system employs a pipeline approach that in- cludes an automatic speech recognition system (ASR) and a machine translation (MT) system. The paper is organized as follows: Section 2 presents the ASR system used, along with a description of conventional, end-to-end, and fusion systems. Section 3 focuses on the MT system in which we describe preprocessing, data augmenta- tion, noisy input translation, ensembling, and reranking com- ponents in detail. We present our concluding remarks in Sec- tion 4. 2. Automatic Speech Recognition In a pipeline-based speech translation system, ASR is the most front-end module. In order to get reliable transla- tion of quality, it is critical to obtain ASR transcription as accurate as possible. To start, we build several con- ventional pipeline-based ASR systems using deep neural network/hidden Markov model (DNN/HMM) framework. In the DNN-HMM framework, we employ several DNNs * Equal contribution with different structures including fully-connected DNN (FDNN), time-delay deep neural network (TDNN) [1, 2], and latency-controlled bidirectional long short-term memory (BLSTM) [3]. Our final goal is to build end-to-end speech translation system, that is, we need to simplify the model-building pro- cess of conventional pipeline-based ASR system by con- structing complicated modules with a single DNN architec- ture or in a data-driven learning method. Thus we also em- ploy an end-to-end ASR system which is based on a hy- brid connectionist temporal classification (CTC) [4, 5] and attention based encoder-decoder [6] architecture. Finally, we combine the output of different ASR systems to boost the final ASR performance. During the development of our system, all acoustic mod- els are trained on TED dataset together with the training datasets provided by the organizer. We noticed that the or- ganizer’s segmentation of the talks/lectures is not quite ac- curate, e.g. some sentences are not properly split. There- fore, we employ our own model-based voice activity …

Content truncated to 500 words