We are excited to report a significant step forward in our ongoing journey to bring innovation to the operational environment of Air Traffic Control (ATC). In our collaboration (Link) with Skyguide, the air navigation service provider of Switzerland, we (DLR) have deployed a demonstrator for Automatic Speech Recognition and Understanding (ASRU) in the live ATC environment of Skyguide.
Since February 19th, our ASRU prototype is officially running in the Dübendorf (Zurich) operations room. Deployed on a non-operational working position in “shadow mode”, the system is safely processing operational data. It operates continuously on the live radio stream and provides immediate transcriptions as well as an automatic understanding of the issued Air Traffic Controller clearances to enable downstream applications such as automatic maintenance for aircraft radar labels in the future. Key features include:
- Frequency Selection: The system is adaptable and can be easily configured for Tower (TWR), Approach (APP), and Area Control Center (ACC).
- Smart PTT-Based Filtering: To ensure accuracy and relevance, the system uses Push-To-Talk (PTT) based filtering. This allows it to accurately transcribe Air Traffic Controller communications. The next iteration will include transcription of pilot utterances as well.
- Enhanced Callsign Recognition: By dynamically feeding the ASRU system with a list of callsigns correlated directly from the live surveillance picture, the prototype achieves a callsign recognition rate of 98%.
The Trial Phase
The current prototype will remain available in the Dübendorf OPS room for a comprehensive trial period running until the end of June 2026. During this phase qualitative feedback and direct performance insights from the Air Traffic Controllers themselves will be collected. The data and hands-on experience gathered during this trial will help in evaluating the real-world maturity and suitability of ASRU technology going forward.
This deployment marks a major step forward in exploring how modern ASRU technologies can be practically applied to support the demanding operational needs of Air Traffic Controllers.
Additional Information
Although the current prototype was not specifically adapted for Zurich airspace, i.e. the ASRU AI models have only seen a very limited amount of training data from the Zurich area. With only round about one hour of voice recordings from Zurich a callsign recognition rate of 98% is achieved. Locally used terms such as waypoints like KOLUL or airline names like Edelweiss or Itarrow are the challenge as they have not been part of the training process for the ASRU AI models. Nevertheless, our ASRU prototype is often capable of extracting the correct ATC clearance, even if the transcript might contain errors. Through contextual knowledge the system automatically figures out to map wrongly recognized words like “edlwiz” to the actual spoken airline “Edelweiss” or the misrecognition “Koll” to the waypoint “KOLUL” etc. The maintenance of the prototype, i.e. the regular update with new airline or waypoint names will be addressed within the SELF-MADE-ATC project, which also investigates and develops tools to enable Air Navigation Surface Providers such as Skyguide to maintain the ASRU prototype by themselves, i.e. without the permanent support of a third party like an ATM supplier.