Clicking time of Air Traffic Controllers (ATCos) reduced from 12,800 seconds down to 405 seconds by using the automatic speech recognition and understanding (ASRU) architecture from the HAAWAII project.
Between the 14th of September and 3rd of November 2022 the DLR conducted, together with 12 ATCos from Austro Control, validation trials for the PROSA project to quantify the benefits of automatic speech recognition and understanding on ATCos’ workload and flight safety.
The following picture shows the participants of the last trials performed on November the 3rd. On this day also, external visitors and interested stakeholders were participating to get an impression of the developed prototype. Validation trials and open day were in the context of exercise 002 for solution 96 of the PROSA project being support by SESAR Industrial Research (funding number 874464).
Each ATCo participated in four exercises, two baseline runs without ASRU support and two solution runs with support of ASRU. Both the solution runs and the baseline runs consisted of a medium and a heavy scenario. The medium scenario contained round about 30 arrivals per hour and the heavy scenario 42 arrivals per hour.
In the baseline scenario the ATCos were asked to manually enter all clearances into the aircraft radar labels, i.e. entering the altitude value of a DESCEND clearance, the speed value of a REDUCE clearance, the cleared heading or waypoint and so on.
The picture below shows the radar display with many radar labels. In the bottom right corner the ATCo is just inputting a value into the radar label of the aircraft AUA774X.
The evaluation is still in progress. Nevertheless, some very clear results are already available. The total simulation time overall runs were 101,400 seconds, i.e. all ATCos together spent 50,700 seconds within the baseline scenarios and also the same amount in the solution scenarios. The ATCos spent in the baseline runs roughly 12,800 seconds for maintaining the clearance information within the aircraft radar labels. We measured the time whenever the ATCO opened a menu to enter a clearance until the desired value was selected. More than 5,900 commands were manually entered in the baseline runs, i.e. an input roughly required 2.2 seconds.
In the solution runs the ATCos only needed to enter a value, when the ASRU failed to recognize the given clearance, i.e. either nothing was recognized or a wrong value or command type were recognized.
The following picture shows the contents of the callsign LOT229, when the heading value 310 was just recognized (purple value) by ASRU.
In all 20 solution runs together only 154 commands were manually entered. The ATCos needed 405 seconds in total for these corrections, i.e. roughly 2.6 seconds. The results are very, very clear: In the baseline scenarios not supported by ASRU the ATCos spent more than 25% of the simulation time with the radar label maintenance. In the solution scenarios supported by ASRU the ATCo’s needed less than 1% of the time for this task. Even if we assume that mechanisms for manually entering the commands in the HMI dramatically improve, an input time of at least one second per command will still remain. That means, the ATCos would still need 12% of their time for the radar label maintenance instead of just 1% when ASRU support is available.
12 to 25 times more additional mental capacity of the ATCo is principle available when ASRU support is available. Therefore, the next steps will be the evaluation of the average flight time per aircraft in the baseline and solution scenarios and, how many given commands were not entered correctly into the radar labels, when ASRU is used and even more important, if also wrong label entries exist, when the ATCo enters all commands manually, i.e. when he has the additional workload with the radar label maintenance task.