Skip to main content

Readback Error Detection by Automatic Speech Recognition to Increase ATM Safety

Paper ID

ATM-2021-007

Conference

USA/Europe ATM R&D Seminar

Year

2021

Theme

Human Factors

Project Name

SESAR 2020 ER4 project HAAWAII, SESAR 2020 IR Wave 2 project PJ10-W2 PROSA

Keywords:

Air Traffic Control (ATC), Automatic Speech Recognition (ASR), Readback Error Detection

Authors

Hartmut Helmke, Matthias Kleinert, Shruthi Shetty, Oliver Ohneiser, Heiko Ehr, Hörður Arilíusson, Teodor S. Simiganoschi, Amrutha Prasad, Petr Motlicek, Karel Veselý, Karel Ondřej, Pavel Smrz, Julia Harfmann and Christian Windisch

DOI

Project Number

874464

Project Number

884287

Abstract

One of the crucial tasks of an air traffic controller (ATCo) is to evaluate pilot readbacks and to react in case of errors. Undetected readback errors, when not corrected by ATCos, can have a dramatic impact on air traffic management (ATM) safety. Although they seldomly occur, the benefits of even one prevented incident due to automatic readback error detection justifies the efforts. This, however, requires highly reliable detections, which is beyond the performance of currently available automatic speech recognition implementations. The HAAWAII project aims to achieve false alarm rates below 10% and readback error detection rates better than 50%. After performing a preliminary analysis by comparing ATCo utterances with pilot readbacks on word level, this approach proves to be very ineffective. Callsigns are abbreviated or not even pronounced, altitude and speed units are often not used, for example “nineteen eight” is the same as “one one nine decimal eight”. Therefore, the presented approach transforms recognized word sequences into so-called ATC concepts, as agreed with the ontology of the SESAR project PJ.16-04. Detecting readback errors on concept level is more reliable and robust as it also considers different forms of conveying the same semantic messages and is also more tolerant to partially misrecognized words. Nevertheless, a good recognition rate on word level is essential to correctly transform words into concepts, which will be achieved by integrating voice data from ATCo utterances and pilot readbacks with context information such as data concerning radar, flight plans, and weather. This paper presents relevant use cases, the ontology-based algorithm, and initial results regarding callsign recognition accuracy for automatic readback error detection purposes.