ATCO2 - Automatic collection and processing of voice data from air-traffic communications
Project Period: 1. 11. 2019 - 28. 2. 2022
Project Type: grant
Agency: European Comission EU
Program: Horizon 2020
air-traffic management, automatic speech recognition, signal processing, legal and ethical framework
Developing machine learning solutions for air-traffic control applications is a challenging task. Besides an expert knowledge, large amount of data for robust performance as well as for validation and verification is typically required. If funded, ATCO2 will deliver a unique platform enabling to collect, store, pre-process and share voice communications data recorded from real world air-traffic control data. The project aims at accessing data from two sources: (a) from certified ADS-B datalinks aligned with a surveillance technology, and (b) directly from air-traffic controllers offered to the project by several air navigation service providers. The technical development will be centred around the ATCO2 platform, built on an existing and extensively used solution of opensky-network partner, ensuring sustainability of the platform after the end of the project. Current platform collects periodically broadcasted aircraft information through a network of ADS-B receivers operated around the globe, further stored at a server. In ATCO2, existing platform will be extended to allow collection, storage and pre-processing of voice communications, and time/position aligned with other aircraft information. Unlike previous works, we will target both channels, i.e. spoken commands issued by air-traffic controllers, and confirmation provided by pilots. In addition to broadcasted data, ATCO2 will have an access to voice recordings from air navigation service providers, namely Austrocontrol. This data will simulate other source of speech recordings (specifically archives), complementing real-time voice communication. The ATCO2 platform will be enhanced by the latest speech pre-processing and machine learning technologies, mostly based on deep learning. Besides automatic segmentation (e.g. er speaker, accent, specific command), robust automatic speech recognition system will be implemented and integrated through RESTful API allowing to automatically transcribe voice communications.