A stochastic speech understanding method to generate interlingual representations

The purpose of this paper is to create an understanding system that can reduce the recognition error that prevents semantic analysis, and that is robust to speech input. A method is proposed in which semantic analysis and speech recognition are integrated on the basis of a statistical model, and the procedure is optimized as a high-level recognition process from speech to semantics. As a method of realizing the idea effectively and efficiently, search of the word graph by means of a decision tree is proposed. Using the proposed method, an identification experiment is performed for the semantics represented by the intermediate language. It is a problem of segmentation of utterances into an average of 1.6 semantic units, and the assignment of the semantic units to 7600 classes characterized by combinations of utterance action and concept. In the speech recognition experiment the understanding error is improved by 8.5% compared to the case in which the 1-best recognition candidate is input.