Lexical out-of-vocabulary models for one-stage speech interpretation

We present an approach to explicit, statistical, lexical-level out-of-vocabulary (OOV) word modeling for direct integration into the search space of a one-stage speech interpretation system. For this purpose, a generic pronunciation model for unknown words is derived from large pronunciation lexica and, optionally, word frequency knowledge. Known statistical language modeling (LM) methods are utilized to estimate different phoneme LM and apply different smoothing techniques. The resulting OOV word models are integrated with the hierarchical language model of our uniform modeling framework by declaring semantically irrelevant parts of the training utterances as unknown. Experiments were conducted with two different OOV training lexica on an airport information dialogue application, evaluating the results with both in-vocabulary (IV) and OOVrelated metrics. Results for various OOV model configurations are presented, showing that OOV detection rates of 60-70% can be achieved with 1-2% falsely accepted IV words, simultaneously improving accuracy on the semantic representation.