KIT’s IWSLT 2018 SLT Translation System

KIT’s IWSLT 2018 SLT Translation System Matthias Sperber, Ngoc Quan Pham, Thai Son Nguyen, Jan Niehues, Markus M¨uller, Thanh-Le Ha, Sebastian St¨uker, Alex Waibel Institute for Antrophomaics and Robotics, Karlsruhe Institut of Technology firstname.lastname@kit.edu Abstract This paper describes KIT’s submission to the IWSLT 2018 Translation task. We describe a system participating in the baseline condition and a system participating in the end-to-end condition. The baseline system is a cascade of an ASR system, a system to segment the ASR output and a neural machine translation system. We investigate the com- bination of different ASR systems. For the segmentation and machine translation components, we focused on transformer- based architectures. 1. Introduction The Karlsruhe Institute of Technology participated in the IWSLT 2018 Evaluation Campaign with systems for English→German Speech Translation task. We submitted system to both conditions: the baseline condition and the end-to-end condition. The submission to the baseline condition is based on the cascaded approach described in [1]. In this evaluation cam- paign, we investigated the combination of different ASR sys- tems. Furthermore, we investigated the use of transformer- based models. This paper is structured as follows. In Section 2, we de- scribe different speech recognition systems we employed in the campaign and how we combined them. Afterwards, we give a detailed description of the segmentation approach in Section 3 and the machine translation system in Section 4. Finally, in Section 5 we describe the end-to-end speech trans- lation model. At the end of the paper we report the results and finish with a conclusion. 2. Speech Recognition In this year’s evaluation, we built three different types of au- tomatic speech recognition systems. All the systems were trained using the data from the TED-LIUM Corpus version 2 [2]. 2.1. Hybrid Model Different from previous years, this year we built only one single HMM-based hybrid model for the speech recogni- tion task. The hybrid acoustic modelling is constructed by stacking 5 LSTMs layers of 320 units, a projection layer of 200 units and a softmax layer to classify 8000 context- dependence phone (CD-Phone) states. As traditional ap- proach, we used Viterbi forced alignment to provide CD- Phone state labels for the training data and the acoustic model was trained using only cross-entropy loss function. For model training, we use SGD with an initial learning rate of 0.004 for 8 epochs and degrade it with a factor of 0.8 for other 8 epochs. We use a momentum term of 0.9 while dropout is set to 10%. Only 40 features of Mel-filterbank coefficients are fed into the LSTMs network every timestep, we did not employ any further speaker adaptation features. After the model was successfully trained, we performed the traditional beam search decoding with the employment of the 4-gram language model. We used Janus Recogni- tion Toolkit (JRTK) [3] as the decoding framework while the language model is built by Cantab research group [4] from WMT data. 2.2. CTC Model Our CTC-based [5] ASR model is similar to the system de- scribed by [6]. …

Content truncated to 500 words