Shinya Kiriyama

Development and evaluation of a spoken dialogue system for academic document retrieval with a focus on reply generation

A spoken dialogue system for retrieving academic documents was constructed and evaluated. In order to realize speech reply with its prosodic features properly controlled to express dialogue focus, a concept-to-speech conversion scheme was developed where the reply sentence concept was finally converted step-by-step to phone and prosodic symbol strings to be fed to a speech synthesizer, instead of simply using a text-to-speech conversion scheme as done in other conventional spoken dialogue systems. The main research focuses were on the following three points: (1) generating a reply sentence easily understood by users and unambiguous to users by controlling the amount of information included in it: deleting redundant information and adding necessary information; (2) realizing functions necessary for increasing system usability; (3) generating a speech reply with natural-sounding prosody, with the dialogue focuses correctly conveyed. Evaluations of the original system, formerly developed, gave us guidelines to realize points (1) and (2). Based on them, a revised system was constructed and evaluated, showing the validity of its new functions. It was also shown that speech replies with prosodic features representing dialogue focuses could bring about clearer information transmission.