ISSUES IN LARGE-VOCABULARY INTERACTIVE SPEECH SYSTEMS

Interactive telephone access to large application databases demands re liable large-vocabulary speech recognition. Current speech recognition performance is limited, but may be boosted by the inclusion of databa se and application information in the recognition process. The use of corroborative alphanumeric information such as spelling can also give a considerable boost to application performance. Accurate models for t he relationships between database contents, the vocabulary used to des cribe these items, and the spelling and pronunciation of this vocabula ry are very important. When these are packaged with careful dialogue d esign, they can make large-vocabulary applications feasible. An exampl e corporate telephone directory system is presented illustrating all o f the above features.