Third Workshop on Spoken Language Technologies for Under-resourced Languages
Cape Town, South Africa
We developed a large vocabulary continuous speech recognition system(LVCSR) for Mongolian language. It is the first LVCSR system of Khalkha dialect in Mongolia. Firstly, we created Mongolian speech corpus for acoustic model and it contains over 6000 utterances in total recorded from 700 different sentences spoken by 40 male speakers, and then we created monophone and triphone based HMMs. Secondary, phoneme, morphone and word based n-gram language models were prepared by using 6 million words in a text corpus. Finally, we conducted continuous speech recognition experiments and obtained the phoneme correct rates of 56% and 67% by using monophone HMMs and triphone HMMs, respectively. We also obtained the word correct rates of 63% and 68% by using monophone HMMs & word based trigram and triphone HMMs & word based trigram, respectively.
Index Terms: Large vocabulary continuous speech recognition (LVCSR), Mongolian, Khalkha, Morpheme
Bibliographic reference. Nakagawa, Seiichi / Turmunkh, Erdenebat / Kibishi, Hiroshi / Ohta, Kengo / Fujii, Yasuhisa / Tsuchiya, Masatoshi / Yamamoto, Kazumasa (2012): "Development of large vocabulary continuous speech recognition system for Mongolian language", In SLTU-2012, 19-23.