Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
Successful and flexible establishment of a spoken dialogue system requires the integration of dialogue design, dialogue management and spoken language processing into a common system environment.
The research and development reported in this paper give details from the establishment of such an environment and its necessary development tools and devices as well as preliminary results from testing the integral continuous speech recogniser component on an artificial CAD-application. The application is characterised by a vocabulary containing names of geometrical objects, words by which these objects may be moved, changed in size, coloured etc, such that speech activated manipulation of the objects may be observed on a display. The interactive functionality is guided by reproductive speech output telling the user if, for instance, he/she is asking the system to perform illegal actions as seen from the pragmatic/semantic component of the dialogue system. The continuous speech recogniser part of the spoken language processing module is based on Continuous Hidden Markov Models (CHMM's) of user-configurable speech units, and the search algorithm is a token-passing Viterbi search algorithm constrained by syntactic rules stored in a finite state table (transition network). In addition to this, stub functions are associated with each state in the network, implementing simple conditions and actions and thereby enabling the implementation of more advanced grammars than otherwise available with traditional finite state tables. These stub functions may also be applied in handling of non-static conditions in the recognition process, such as pragmatic constraints. The artificial CAD application has been used to test the system, and the performance is reported for a traditional grammar both with and without stub augmentation. Results from these experiments show that improvements of the recognition rate may be obtained by implementing pragmatic and semantic constraints as directly associated stub functions within the states of the network.
Bibliographic reference. Lindberg, Boerge / Andersen, Bjarne / Baekgaard, Anders / Broendsted, Tom / Dalsgaard, Paul / Kristiansen, Jan (1992): "An integrated dialogue design and continuous speech recognition system environment", In ICSLP-1992, 1653-1656.