TRIPS Component: Speech Recognition

TRIPS employs user-independent speech recognition using Sphinx-II, a general-purpose speech recognizer from CMU. We use Sphinx' generic acoustic models and create our own language models based on the task domain we are implementing. The result of recognition is a set of word hypotheses which is sent to the parser.

Building a language model for a new domain requires a large corpus in the domain, which can be time-consuming and expensive. We have successfully used two techniques to circumvent this. First, we generate an artificial corpus based on a hand-coded, task specific CFG. Second, we use a class-based approach to "port" out-of-domain corpora to our new domain. This allows us to effectively re-use other problem-solving dialogues that we have previously collected.

Message Traffic

• to Parser

Current Researchers

• Lucian Galescu
• Paul McCarty

Relevant Papers

• Lucian Galescu, Eric K. Ringger, and James F. Allen. "Rapid Language Model Development for New Task Domains." Proceedings of the ELRA First International Conference on Language Resources and Evaluation (LREC). Granada, Spain. May 1998. ( gzipped postscript)