The architecture of the system is shown in Figure 1. The modules communicate asynchronously by message passing; hence, in principle all of them could run in parallel in different processes. In the current implementation, there are four processes, which handle speech recognition, speech synthesis, database access and everything else, respectively.
The speech recognizer is a Swedish-language version of the SRI Decipher
system [Murveit et al.
1993], developed by SRI International and Telia
Research. It sends an N-best speech hypothesis list to the two language
processors: the Core Language Engine (deep analysis) and the Robust
Parser (shallow analysis), further described in
Section 
. The language processors each
send their analyses to the dialogue manager (DM). After each
system turn, the DM updates the language processors with limited
information about the state of the discourse: the most recent question
(if any) posed by the system, and the types of objects that are salient
at the current point in the dialogue.
The DM uses a two-stage heuristic selection process to advance the
dialogue. First, each input analysis is categorized as a move of a
certain type, and an appropriate response to that move is
selected. References are resolved and contextual information is also
added, resulting in a further multiplication of possible moves and
responses. Secondly, the relative utility of the various responses is
judged, and the most productive response move is chosen. The dialogue
manager is further described in Section .
The generator produces the surface string representing the actual utterance, using a simple template-based approach. The surface string is then turned into speech by Telia Research's synthesizer LIPHON.
In the current system, the database agent contains a web client in order to retrieve data from the Travellink database. All query results are cached in order to shorten the response times as much as possible. However, the response times for most queries would clearly not be acceptable in a commercial system. That inspired us to develop a version that is able to continue the dialogue while database access is in progress (that is, the system might ask about the return leg of a trip, while the database agent is searching for possible trains or flights for the outbound leg).
The system described here is fully implemented and has been permanently installed at the Telia Vision Center in Farsta/Stockholm since November 1998.