NASA NTRS2005
Clarissa, an experimental voice enabled procedure browser that has recently been deployed on the International Space Station, is as far as we know the first spoken dialog system in space. We describe the objectives of the Clarissa project and the system's architecture. In particular, we focus on three key problems: grammar-based speech recognition using the Regulus toolkit; methods for open mic speech recognition; and robust side-effect free dialogue management for handling undos, corrections and confirmations. We first describe the grammar-based recogniser we have build using Regulus, and report experiments where we compare it against a class N-gram recogniser trained off the same 3297 utterance dataset. We obtained a 15% relative improvement in WER and a 37% improvement in semantic error rate. The grammar-based recogniser moreover outperforms the class N-gram version for utterances of all lengths from 1 to 9 words inclusive. The central problem in building an open-mic speech recognition system is being able to distinguish between commands directed at the system, and other material (cross-talk), which should be rejected. Most spoken dialogue systems make the accept/reject decision by applying a threshold to the recognition confidence score. NASA shows how a simple and general method, based on standard approaches to document classification using Support Vector Machines, can give substantially better performance, and report experiments showing a relative reduction in the task-level error rate by about 25% compared to the baseline confidence threshold method. Finally, we describe a general side-effect free dialogue management architecture that we have implemented in Clarissa, which extends the "update semantics'' framework by including task as well as dialogue information in the information state. We show that this enables elegant treatments of several dialogue management problems, including corrections, confirmations, querying of the environment, and regression testing.