English to Japanese spoken language translation system for classroom lectures

This paper presents our attempt to create English automatic speech recognition (ASR) and English to Japanese statistical machine translation system (SMT). We used MIT OpenCourseWare lectures as our test lecture corpus. Wall Street Journal (WSJ) corpus adapted with MIT OpenCourseWare lectures was used as our acoustic model. MIT OpenCourseWare lecture transcriptions were utilized to create our language model. As for the parallel corpus, we used TED Talks and Japanese-English News Article Alignment Data (JENAAD). Our proposed ASR system can achieve 32.1% 0word error rate (WER) and our SMT system can achieve 10.95 BLEU.