Coalescence Type based Confidence Warping for Agglutinative Language Keyword Spotting

In agglutinative languages like Korean, words are formed by joining l affix morphemes to the stem, which leads to high OOV rate in dictionary building. Hence, subword units are usually used as basic language modeling units in Large-Vocabulary Continuous Speech Recognition (LVCSR) or LVCSR based applications such as keyword spotting. In this work, firstly a new word property called coalescence type is introduced, which is defined based on the result of word segmentation process and thus unique for agglutinative languages. A confidence warping approach is then proposed to adjust confidence measure for keyword candidates, with the additional linguistic level information. An evaluation on Korean telephone speech keyword spotting task shows that up to 2% improvement can be obtained in precision, which is significantly better than the baseline system.