Estimation of language models for new spoken language applications

Spoken language interfaces can provide natural communication for many database retrieval tasks. The CMU ATIS system provides an example of accessing airline information using spoken natural language queries. However, a lot of training data is needed to develop a spoken language application. For example, we need training data to generate a language model that can be used by the recognizer to reduce the search space. In this paper, we will address some issues arising from small amount of training data available for a new spoken language application. We are working on a spoken language interface to access information from a library catalogue. The catalogue contains around 13,000 titles, 6000 authors and 19000 subjects. There are more than 20,000 words in the dictionary. The user can seek information about books, authors, subjects, publishers, etc. For example, “I’d like to see books dealing with science fiction by Clarke.” We will describe some language modelling experiments for this task. We will briefly describe a speech interface [5] for a library catalogue. We will also review class-based language models and describe their limitations. Finally, we will present our approach to building statistical language models for new spoken language applications. This is important because a lot of training data is normally needed to generate a language model. However, it is not practical to have or collect a large corpus of data for each new spoken language application.