Module Command Prediction -- Description of Inputs and Outputs

Kurzfassung

This document describes the C++ module CommandPrediction (libCommandPrediction, ModContext_cxx), which forecasts the most probable air traffic control (ATC) concepts such as callsigns and commands that an air traffic controller (ATCo) will most likely utter in radiotelephony with aircraft crews during the next minutes. The output of the CommandPrediction, also called HypothesesGenerator, supports other modules of the automatic speech recognition and understanding (ASRU) system in improving their output and in checking the plausibility of their output. For example, if the Speech-to-Text module or the Text-to-ATC-concepts [1] module contain “speed bird two seven victor” or “BAW27V” respectively, but this callsign is not part of the set of callsigns from CommandPrediction, it can be discarded to reduce potential errors. In a similar way, the ASRU output of a TAXI-command could be discarded if the set of predicted commands only contain DESCEND and HEADING commands due to an aircraft being for example, in flight level three hundred. The most important use case of this module is, however, if the output of Speechto-Text module contains wrong words, e.g. “speed bird three seven victor” instead of “speed bird two seven victor”, then the wrong word can often be corrected. The format of prediction contents (as all other ASRU modules) follow the defined ontology for the annotation of ATC commands [2]. The format for the exchange of prediction content is JSON. The use of CommandPrediction is optional, but usually helps to significantly reduce the ATC concept recognition error rate and increase the command recognition rate of the ASRU system overall (see ASRU components in Figure 1-1). The CommandPrediction module receives surveillance data (radar data and flight plans) as well as meteorological data if available.