Word class modeling for speech recognition with out-of-taskwords using a hierarchical language model

Out-of-vocabulary (OOV) problems are frequently seen when adapting a language model to another task where there are some observed word classes but few individual words, such as names, places and other proper nouns. Simple task adaptation cannot handle this problem properly. In this paper, for task dependent OOV words in the noun category, we adopt a hierarchical language model. In this modeling, the lower class model expressing word phonotactics does not require any additional task dependent corpora for training. It can be trained independent of the upper class model of conventional word class N-grams, as the proposed hierarchical model clearly separates Inter-word characteristics and Intra-word characteristics. This independent-layered training capability makes it possible to apply this model to general vocabularies and tasks in combination with conventional language model adaptation techniques. Speech recognition experiments showed a 19-point increase in word accuracy (from 54% to 73%) in the with-OOV sentences, and comparable accuracy (85%) in the without-OOV sentences, compared with a conventional adapted model. This improvement corresponds to the performance when all OOVs are ideally registered in a dictionary.