Incremental Semantic Models for Continuous Context-Sensitive Speech Recognition

Abstract:

Context-sensitive speech recognizers use environment or discourse information to influence language model probabilities used in speech decoding. This is usually done by switching language models between utterances. This paper explores the use of a continuously context-sensitive language model that uses incremental interpretation to update context at every time step in decoding. Because it only explores the world model incrementally, this semantic model does not need to be pre-computed, raising the possibility of representing continuouslyvariable concepts as semantic referents (such as time points and measurements, or real numbers themselves), and supporting dynamic reasoning about consequences of actions during speech decoding.