Dialogues in Air Traffic Control

We have taken an off-the-shelf, commercial continuous speech recogniser and conducted evaluations for the domain of Air Traffic Control. The language of this domain proved to be quite unrestricted, contrary to our initial intuitions. Our experiments show that constraints typically used by speech recognisers do not provide accurate enough results and need to be augmented with other knowledge sources and higher levels of linguistics in order to prove useful. We used three syntaxes based on a corpus of transmissions between the ATC and pilots in order to reflect differing levels of "linguistic" knowledge. Initial experiments demonstrate the benefit of a fully constrained context-free semantic grammar. Further experiments empirically show the benefit to recognition accuracy of using some form of dialogue management system to control the flow of discourse. A corpus-based statistical clustering approach to the segmentation of a dialogue into discourse segments is briefly discussed.