Categorical understanding using statistical ngram models

In this paper, the speech understanding problem in the context of a spoken dialog system is formalized in a maximum likelihood framework. Word and dialog-state n-grams are used for building categorical understanding and dialog models, respectively. Acoustic con dence scores are incorporated in the understanding formulation. Problems due to data sparseness and out-of-vocabulary words are discussed. Incorporating dialog models reduces relative understanding error rate by 1525%, while acoustic con dence scores achieve a further 10% error reduction for a computer gam-