 |
Error Handling in Spoken Dialogue SystemsAugust 28-31, 2003 |
 |
|
Error Handling in Spoken Dialogue SystemsAugust 28-31, 2003 |
|
The present contribution adresses the design of pragmatic solutions for various problems occuring within dialogue-based vocal systems using low performance speech recognition engines (SREs), for example in situations where the used speech recognition engine is not fully adapted to the specific application, or the data necessary for training reliable acoustic and language models is not available. To accomodate the use of a low performance SRE, the following design principles are used to guide the conception of the dialogue model: (1) adopt a (limited) mixed initiative dialogue management strategy to improve flexibility of use; (2) avoid repetitions in the dialogue flow; (3) integrate in the dialogue management strategy mecanisms for recovering from specific dialogue repair situations such as request for help, for repetition, miscommunication, ...; (4) minimize the duration of the dialogues, i.e. aim at dialogues providing the user with the relevant information in a minimal number of turns; (5) provide the user with adequate feedback information about the state of the dialogue and the recognized pieces of information; and (6) filter out as much conflicting data as possible.
The structure of the contribution is the following: we first detail the context in which our dialogue model was designed; then we describe the solutions that have been proposed to implement the above mentioned design principles. Next, from the final evaluation of the system, we derive some insights on the impact of the selected solutions on the user perception of the system. All the proposed solutions were designed, implemented and evaluated in the framework of the InfoVox project.
Bibliographic reference. Seydoux, Florian / Trutnev, Alex / Rajman, Martin (2003): "Dialogue management with weak speech recognition: a pragmatic approach", In EHSD-2003, 133-138.