Data-Driven Determination of Appropriate Dictionary Units for Korean LVCSR

Abstract

This paper describes the design of our Korean large vocabulary speech recognition system using the multilingual dictation database GlobalPhone. Defining appropriate dictionary units for this purpose is not a trivial task since using word phrases (eojeols) gives very high OOV-rates, above 30%, whereas using syllable units results in high confusabilities and a very limited scope of standard language models. We investigate a data-driven approach which overcomes these limitations. The results show that the data-driven approach reduces the OOV-rate to below 1 % and significantly outperforms the syllable based approach according to phone and syllable accuracy giving 79.4 % and 69.3 % accuracy respectively. For our best system we present lattice based accuracies achieving 95.0 % syllable accuracy and 82.7 % eojeol accuracy.