6 Discussion and Future Work

Computation of the probability of initial sub-string generation by stochastic context free grammars. A-posteriori considerations on this rst phase of acquisitions have been positive. First of all, nearly every subject believed the system was true, at least as concerns the speech recognition module. This result is particularly satisfying as about half of the subjects were researchers, many of them working in the realm of AI. Second, the collected material has turned out to be mostly usable for the design of future prototypes of the related systems. Moreover, sentences also present great linguistical diierences with respect to the simulated systems. Commands to the telecontrol station are usually expressed by concatenations of short and simple sentences (with or without using explicit conjunctions). In order to build a spoken language interface for the mobile robot, a regular grammar covering the navigation shell sentences is being designed, to be included into a Hidden Markov Models based speech recognizer. Moreover, a command interpretation module to be interfaced with the navigation shell of the actual robot is also being developed. Finally, the collected material seems to be suucient to carry out both these modules. As shown in Section 4, concierge simulations provided sentences with more structural and lexical variability. These transcriptions are currently being syntactically and semantically analyzed with respect to the grammar, lexicon, and knowledge base of the actual system. As a result of such a work, we expect to get useful indications for the design of a second phase of acquisitions; the material collected in these two phases will be employed in the construction of the planned spoken language interface. 7 Acknowledgements The authors wish to thank the colleagues at IRST who contributed to the design and the development of the experiment. In particular, the Computer Vision Group contributed in the design of the Mobile Robot Telecontrol Station simulation , the Natural Language Processing Group to the Concierge one. We are especially grateful to the other members of the Speech Group for the overall support in carrying out the simulations and transcriptions. Finally, we are in debt with Alberto Lavelli for revising this manuscript. To evaluate the acoustic isolation of the subject room, measurements of the average (over time) background noise level were performed. Before the acoustic isolation values between 39 and 48 dBA (depending on the hours of the day in which the measurements were done) were measured, while the complete isolation of …