This thesis focuses primarily on constructing voice-only pedestrian
guidance systems using spatial database techniques. In the process of
doing this we first explored how to use authoring tools to build
natural language interfaces over large databases. Specifically we
built a natural language interface over the MusicBrainz
database of 1.5GB and confronted the resulting scalability issues. We
then explored vague querying, specifically spatial queries using
Assuming `language as a set of conventions', we proposed an approach
for handling vagueness by defining contexts that are compiled to crisp
SQL view definitions.
In our recent work, as partners in the Spacebook
project (http://www.spacebook-project.eu), we have focused on
how to build reliable, scalable and extensible
text-to-speech (TTS) based navigation systems for pedestrians.
Technical aspects we have worked on include building the system
Janus (http://janus-system.eu), with sensor reports and
bidirectional voice channels. Experimental work has been mostly focused
on measuring accuracies and latencies with available hardware. We have
also, very recently, started human usability experiments. Our
theoretical work has been in defining models for how a system can
interact with pedestrians over high latency data links with poor GPS
quality using prediction of pedestrian positions and scheduling
utterances to mask latencies. To allow for scalable deployment, we
have only used standard smart phones and inexpensive servers.