The 2017 KIT IWSLT Speech-to-Text Systems for English and German

The 2017 KIT IWSLT Speech-to-Text Systems for English and German Thai-Son Nguyen, Markus M¨uller, Matthias Sperber, Thomas Zenkel, Sebastian St¨uker and Alex Waibel Institute for Anthropomatics, Karlsruhe Institute of Technology Karlsruhe, Germany {first.lastname}@kit.edu Abstract This paper describes our German and English Speech- to-Text (STT) systems for the 2017 IWSLT evaluation cam- paign. The campaign focuses on the transcription of un- segmented lecture talks. Our setup includes systems using both the Janus and Kaldi frameworks. We combined the out- puts using both ROVER [1] and confusion network combina- tion (CNC) [2] to achieve a good overall performance. The individual subsystems are built by using different speaker- adaptive feature combination (e.g., lMEL with i-vector or bottleneck speaker vector), acoustic models (GMM or DNN) and speaker adaptation (MLLR or fMLLR). Decoding is per- formed in two stages, where the GMM and DNN systems are adapted on the combination of the first stage outputs using MLLR, and fMLLR. The combination setup produces a final hypothesis that has a significantly lower WER than any of the individual sub- systems. For the English lecture task, our best combination system has a WER of 8.3% on the tst2015 development set while our other combinations gained 25.7% WER for Ger- man lecture tasks. 1. Introduction For many years now, the International Workshop on Spo- ken Language Translation (IWSLT) offers a comprehensive evaluation campaign on spoken language translation. The evaluation is organized in different evaluation tracks cover- ing automatic speech recognition (ASR), machine translation (MT), and the full-fledged combination of the two of them into speech translation systems (SLT). Different from previ- ous years, this year’s installment mostly consists of real-life lectures e.g., real university lectures or talks at real symposia. The goal of the ASR track is the automatic transcription of fully unsegmented lectures. The quality of the resulting transcriptions is measured in word error rate (WER). This system paper describes our English and German ASR setups with which we participated in the lecture ASR tracks of the 2017 IWSLT evaluation campaign. Similar to previous years’ evaluation [3], we used the Janus Recogni- tion Toolkit (JRTk) [4] which features the IBIS single-pass decoder [5] to build several complementary subsystems and combined them with an additional system developed with the Kaldi toolkit [6]. Our Janus-based systems employ different speaker-adaptive features, acoustic models or speaker adap- tation techniques. While the Kaldi-based system applies the same adaptation techniques but employs sequence training and big n-gram language models for rescoring. The rest of this paper is structured as follows. Section 2 describes the data that our system was trained and tested on. This is followed by Section 3 which provides a description of the acoustic front-ends used in our system and Section 7 which describes our segmentation setup. An overview of the techniques used to build our acoustic models is given in Sec- tion 5. We describe the language model used for this evalua- tion in Section 6. Our decoding strategy and results are then presented in Sections 8 and 9. We …

Content truncated to 500 words